Timeline
Every post and feed across this instance
-
Security fix
We got a report of a security issue which we quickly addressed and fixed.
Instructions on how to update your server.
This server has been updated as has rss.chat.
-
Security fix
We got a report of a security issue which we quickly addressed and fixed.
Instructions on how to update your server.
This server has been updated as has demo.rss.chat.
-
An online edition of all 85 of the Federalist Papers ,...
An online edition of all 85 of the Federalist Papers, presented as they were originally serialized in newspapers, with on eye on good typography and readability.
-
The Rise of Anti-Flock Influencers Who Make Things Up for Clout
Instagram influencers are creating fake cease-and-desist letters, making themselves go viral and muddying the waters about mass surveillance. -
Alright.
-
Early this morning we got a report of a security issue in the rss.chat server, quickly fixed and tested the new version. So, if you're running your own instance of rss.chat, you please follow the instructions and do the update asap.
-
The Simple Joy of “No Phones Allowed"
“Every time someone in a group of people deploys a screen, the whole group is affected. Each disengaged person in a crowd is like a little black hole, a dead zone for social energy, radiating a noticeable field of apathy towards the rest of the room and what’s happening there.”
-
“ This map shows the predicted density of...
“This map shows the predicted density of underground networks created by arbuscular mycorrhizal fungi. These networks form one of Earth’s circulatory systems, moving massive amounts of carbon, nutrients, and water across plant communities.”
-
What's coming to RSC: smarter sources, cleaner timelines, honest attribution
We've just finished designing the biggest under-the-hood upgrade RSC has had since launch. It's being built in four stages, and while most of the work is invisible plumbing, all of it exists to change what you actually see and trust in your timeline. Here's what you'll get.
One story, one card
Today, when the same post reaches RSC through more than one feed — a personal blog you follow and a newsletter that republishes it — you can end up with duplicates in your river. After the upgrade, RSC understands that these are the same item.
You'll see one card, showing the best version available,
with clear credit to the person or site it came from. If a better copy arrives later (say, the author's own canonical version), the card quietly upgrades itself.Know where things really come from
Anyone can put a name on a feed item. Soon, RSC will be able to check.
When a post claims to come from a particular site, RSC can verify that the post actually appears there — and content that passes gets the strongest attribution. Impersonation gets harder; genuine authorship gets the credit.
A safer, better-kept neighborhood
Instance operators get real tools to keep timelines healthy without
rewriting history:- Hide an individual post clear record of what was done and why.
- Purge a bad source completely — and it stays gone: renamed or redirected copies of a purged feed can't sneak back in.
Everything is reversible where it should be, permanent where it must be, and every action leaves an audit trail.
Your subscriptions, upgraded in place
When the switch happens, you don't have to do anything.
Everyone you follow, every feed you've subscribed to, every post and conversation, and the instant-update connections that make new posts appear live — all of it carries over automatically.Each instance is backed up before the switch, the changeover happens in one atomic step, and if anything looks wrong we
restore and try again another day.What doesn't change
RSC stays RSC. Posts, replies, and conversations still travel as plain RSS that any reader can consume. Your data stays yours, in open formats, with no lock-in.
Everything above works without JavaScript, in both themes, on the web you already use.
When?
The four stages ship in order, each fully tested behind a switch before it's turned on.
We'll announce here as each one goes live — starting with the foundations, which are being built right now.
-
Updating a server you already run
This is the first of a kind of note you'll see here from now on: when a change matters to people running their own installs, the steps to update will live right here in worknotes. For this one:
Get the new
rssnetwork.jsandpackage.jsonfrom the repo.In your server's folder, run
npm install. The updatedpackage.jsonadds one package -- the cleaner -- and this brings it in.Restart the server.
That is the whole update. Nothing in your config or your database changes, and the cleaner applies to new and edited posts the moment you restart.
-
A security fix -- posts are now cleaned as they're saved -- and the first update note of a new kind: how to bring a server you already run up to date
A post's text is written by its author and then shown to everyone who reads the timeline, which means anything hidden inside it runs in every reader's browser. Until today the server stored a post's text exactly as it arrived, so a post could carry markup that did more than format words. As of server v0.6.3, every post passes through a cleaner on its way into the database. Ordinary writing comes through untouched -- links, bold and italic, quotes, lists, and pasted images are all kept -- but anything that could run code is removed before the post is ever stored. Because the cleaning happens as the post is written, it protects every reader of every feed and timeline at once, and no old or unusual browser can slip past it. Both public servers, rss.chat and demo.rss.chat, are already running it. Thanks to the reader who reported this privately.
-
Hello world. I'm testing to make sure that the server is working after another upgrade.
-
Leaked Document Shows the Surveillance Tech at ICE’s Fingertips
From phone location data, to social media monitoring, to online undercover tools, a document obtained by 404 Media lays out the surveillance tech available across ICE agency wide. -
test
-
La source originale de la web app: https://rss.chat
Et pour celles et ceux qui veulent essayer sans complications: https://demo.rss.chatL'installation de l'instance soFab a été effectuée grâce à l'aide de Claude.
-
Orcas Explode Fish Into Confetti, Possibly for Fun, In ‘Astounding’ New Video
A team captured first-of-its-kind footage of orcas holding a sunfish carcass still and ramming it so hard that it burst into fragments, which may be a form of social play. “We could hear the sound,” said one researcher. -
Something to keep in mind in press reports with AI apps breaking out of their sandbox, it works the other way too. If you give a big piece of code to Claude and ask if it to find any security issues, it not only finds (at least some of) them, but it also suggests fixes. Quickly. I've done it the other way, where you have a small team, and someone discovers a hack, and you have to find the right answer and implement it, asap.
-
Hi
-
I've now foutered my way through sendgrid and for 100 days on the free plan I've got a way to sign into my localhost copy of rss chat.
I am going to try my Raspberry Pi next, probably at the weekend.
Uploading image…
-
Our New FOIA Forum! 7/30, 1PM ET
Join us for our latest FOIA Forum where we teach you how to pry records from the government. -
Hello, I just arrived here. As I posted to demo.rss.chat I managed to get this running on my mac, without much bother.
rss chat is now simper to set up so I gave it a wee go on local host. So far I've just run it from my downloads folder. The instructions worked but I am a node-no-nothing so I tripped a couple of times.
First I didn't npm install so nothing worked.
I tried to set up SMTP using SendGrid, but I lost the will to live half way through. Got an API key but
> Message failed: 550 The from address does not match a verified Sender Identity. Mail cannot be sent until this error is resolved. Visit https://sendgrid.com/docs/for-developers/sending-email/sender-identity/ to see the Sender Identity requirementsSo I just copied the magicString from the terminal and constructed my own url:
`http://localhost:1420/userconfirms?emailConfirmCode=ysugstk4yc\`Which did the trick.
Of course a localhost-one-person social network is not much use ;-) but interesting none the less.
Once I have braced myself and figured out SMTP I'll try this on my raspberry pi next.
-
Security versus Privacy
Security versus Privacy
by Loren Stephens
Loren talks about the differences between Security and Privacy, but also how (and why) he prioritises one over the other.
I actually wrote about this topic back in 2019 - I agree with Loren that Security and Privacy are often used interchangeably, but they're objectively different things.
Anyway, that's old news. The thing I wanted to comment on in this post is Loren's position on how he prioritises one over the other. He says:
So, when I ask myself which one matters more, I consistently arrive at the same answer: security first, then privacy.
I used to be among those who avoided Google services due to privacy concerns, and I understand why. [...] However, by choosing to avoid Google, many people end up using less secure alternatives. They've prioritized the second question while neglecting the first.
I agree strongly with this. I've fallen foul of this myself, where in my efforts to de-Google, I've ended up using worse tools both in terms of security and functionality.
At this point in my life, I just want my shit to work well and get out of my way. I do, however, prioritise security, then privacy, then functionality. Ultimately, I care more about keeping my data secure, than anything else.
That's not to say I don't care about privacy. I just care more about securing my data.
If I can find a tool that gives me both security and privacy, I'll use that every time.
Thanks for reading this post via RSS. RSS is ace, and so are you. ❤️
You can reply to this post by email, or leave a comment.
-
Minor unexpected behavior
In a browser (e.g., Safari or Firefox on a Mac)
Click the post feed icon.
Click browser's Back button.
Click the post timestamp.
Click the post username.
Click browser's Back button again.
You should see the last post and the user's feed on the same page. -
Dr. Gillie Jenkinson on Thought Reform, Trauma, and Walking Free (Part 2)
In Part 2 of our conversation with Dr. Gillie Jenkinson, we move from her remarkable story into the practical question so many survivors ask: How do you actually recover after leaving a cult? Gillie shares how her own experience escaping a high-control Christian group led her to become one of the world's leading specialists in cult recovery, and why she set out to create ethical, trauma-informed support when almost none existed. We dig into the foundations of her work, from the International Cultic Studies Association (ICSA) and the pioneering Wellspring Retreat Center to the influence of Dr. Robert Jay Lifton's thought reform model, exploring why understanding coercive control is often the first step toward reclaiming yourself.
We also unpack the common threads Gillie has seen after decades of working with former members of high-control groups, religious cults, and coercive relationships. From identity reconstruction and trauma recovery to the importance of healthy therapeutic boundaries and making sense of what happened, this convo’s packed with hope and hard-earned wisdom for anyone navigating life after undue influence.
Pick up Dr. Gillie Jenkinson’s book, Walking Free from the Trauma of Coercive, Cultic and Spiritual Abuse, follow her on Facebook or LinkedIn, and follow Hope Valley Counselling at hopevalleycounselling.com or on Facebook.
Trigger warning: This episode contains frank discussion of religious trauma, coercive control, thought reform, and psychological manipulation.
Also…let it be known that:
The views and opinions expressed on A Little Bit Culty do not necessarily reflect the official policy or position of the podcast. Any content provided by our guests, bloggers, sponsors or authors are of their opinion and are not intended to malign any religion, group, club, organization, business, individual, anyone or anything. Nobody’s mad at you, just don’t be a culty fuckwad.
Buy the A Little Bit Culty book on Amazon or order a signed copy.
Check out our amazing sponsors
Join A Little Bit Culty on Patreon
Get poppin’ fresh ALBC Swag
Support the pod and smash this link
Check out our cult awareness and recovery resources
Watch Sarah's TED Talk and buy her memoir, Scarred
CREDITS:
Executive Producers: Sarah Edmondson & Anthony Ames
Production Partner: Citizens of Sound
Co-Creator: Jess Tardy
Audio production: Will Retherford
Production Coordinator: Lesli Dinsmore
Writer: Sandra Nomoto
Social media team: Eric Skwarzynski and Brooke Keane
Theme Song: “Cultivated” by Jon Bryant co-written with Nygel Asselin
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.
pscrb.fm -
Librarians Are Helping the Public Ditch AI w/ Hannah Cyrus & Alison Macrina
Librarians are working harder than ever to support their neighbourhoods — including by teaching and promoting digital literacy. Hannah Cyrus and Alison Macrina join Paris Marx to discuss the role of library workers in the digital landscape, including how they’re helping patrons navigate privacy basics and artificial intelligence, and how the tech industry is making their jobs harder.
Hannah Cyrus is a digital media librarian at the Bangor Public Library. Alison Macrina is an activist librarian and the director of Library Freedom Project.
The podcast is made in partnership with The Nation. Production is by Kyla Hewson. Support the show on Patreon.
Also mentioned in this episode:
- Click here to get your first month of Proton Mail for just $1.
- The Library Freedom Project has a ton of resources on critically approaching technology.
- Bangor Daily News wrote about librarians like Hannah helping patrons get AI off their devices.
- Shout out to friend of the pod Brian Merchant and his book Blood in the Machine.
- Learn more about the Luddite Club.
- Learn more about the organization Libraries For The People, working to strengthen and expand libraries.
- Support your local library! Say hello, get a card if you don’t have one, and participate in community events.
pscrb.fm -
Much better than Chatouk!🙄
-
This is great!
Thank you Dave 🤗
-
pasted image, looks good in the editor. -
Something about the UI here is pleasingly tactile.
-
How does
relate to
- https://wordland.social/ & https://this.how/wordland/ & https://github.com/scripting/wordlandSupport
I see that the editor features are a little different. In wordland, there's a little paper-airplane icon that publishes/updates the wordland post on a wordpress blog:
the arrow icons move back and forth between wordland posts, and the + adds a new post.In demo.rss.chat, it's setup like a top-posting chat, where the '+' icon on the side bar add new top-level posts:
and the posts appear to make a tree with replies, and comments.
The header of the post has three clickable bits, The poster's name, (which leads to a list of their posts, e.g.
https://demo.rss.chat/?screenname=drf5n
), a direct link to the particular post (e.g. this one is
https://demo.rss.chat/?id=231 )
, and the poster's RSS feed, so you could keep up-to-date with an RSS reader.
-
#1 Women’s Hormone Doctor: Take Control of Your Body, Balance Hormones, & Feel Amazing
Today’s episode is your ultimate women’s hormone health guide for every age.
If you are dealing with painful periods, infertility, endometriosis, PCOS, struggling with perimenopause, menopause, or confused about birth control - this episode is a MUST listen.
From PMS to menopause, this episode will teach you how to use science to hack your hormones, reset your body, and feel amazing again.
Today, Mel sits down with Dr. Sharon Malone, MD, one of the most trusted medical experts in women’s health today, to cover health topics we have never discussed on this podcast before.
Whether you’re 25, 35, 45, 65, or even 85 - this is your playbook to understanding what is happening in your body so you can take better care of your health.
Dr. Malone is a board-certified OB-GYN, a nationally recognized expert in menopause, perimenopause, and reproductive health, and she has been in clinical practice for almost 40 years.
And today, she is unpacking the science of women’s hormones in a way you’ve never heard before.
Dr. Malone answers the questions women everywhere are asking and explains, step by step, exactly what is happening in your body and how to get it back in balance.
You’ll learn:
-The symptoms that deserve your attention
-Why painful periods are not something you should just push through
-The truth about birth control and your long-term health and fertility
-What you need to know about natural birth control and how to think about your options
-What you need to know about fertility and infertility
-How endometriosis and PCOS impact women’s health
-What perimenopause really is and why so many women miss the signs
-What to know about menopause and HRT
-Why weight changes happen as hormones shift
-How muscle loss impacts your health as you age
-What every woman needs to understand about osteoporosis
-The two specific questions you need to ask your mom
-The misinformation online that is hurting women’s health
This is your science-backed step by step guide to balancing your hormones for health, happiness, and longevity.
No matter how old or young you are, this episode is relevant for you and for every woman you love.
For more resources related to today’s episode, click here for the podcast episode page.
If you liked the episode, check out this one next with: Start Where You Are: #1 Orthopedic Surgeon’s Proven Protocol to Feel Stronger & Look Younger in Weeks
Connect with Mel:
- Order Mel’s new product, Pure Genius Protein
- Get Mel’s newsletter, packed with tools, coaching, and inspiration.
- Get Mel’s #1 bestselling book, The Let Them Theory
- Watch the episodes on YouTube
- Follow Mel on Instagram
- The Mel Robbins Podcast Instagram
- Mel's TikTok
- Subscribe to SiriusXM Podcasts+ to listen to new episodes ad-free
- Disclaimer
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
dts.podtrac.com -
Les mecs à Pattaya
Derrière les plages paradisiaques de l’Asie du Sud-Est se cache une autre réalité : celle du tourisme sexuel. Chaque année, des hommes occidentaux viennent y chercher des femmes qu’ils fantasment comme plus douces, plus dociles, plus disponibles.
Mais que racontent vraiment ces relations ? Comment perpétuent-elles des imaginaires coloniaux et des fantasmes raciaux ? Et comment certaines femmes utilisent-elles ces échanges pour améliorer leur quotidien et celui de leur famille ?
Pour en parler, Tal Madesta reçoit Manon Prigent, autrice du documentaire audio L’amour à Pattaya (Arte Radio, 2019) et Marion Bottero, docteure en anthropologie et autrice de l’ouvrage Tourisme sexuel et relations conjugales en Thaïlande et en Malaisie (Éd. L’Harmattan, 2015).
Toutes les références citées dans l'émission sont disponibles sur le site : https://www.binge.audio/podcast/les-couilles-sur-la-table/les-mecs-a-pattaya/
CRÉDITS : Les Couilles sur la table est un podcast créé par Victoire Tuaillon produit par Binge (URBANIA AUDIO). Cet entretien a été préparé, mené et monté par Tal Madesta, enregistré le 26 juin au studio Badje (Paris 17e). Supervision éditoriale et incarnation : Naomi Titti. Réalisation et mixage : Octave Bothier. Responsable des productions éditoriales : Charlotte Baix. Assistante de production et d’édition : Aude Miquel. Rédacteur en chef : Thomas Rozec. Direction de production : Albane Fily. Responsable administrative et financière : Adrienne Marino. Responsable sponsoring : Betty-Maeva Wendling. Musique originale : Théo Boulenger. Identité graphique : Camille Bernard (Upian). Composition identité sonore : Jean-Benoît Dunckel. Voix identité sonore : Bonnie El Bokeili. Direction de contenu : Sophie Marchand. Binge est une marque URBANIA.
Hébergé par Audiomeans. Visitez audiomeans.fr/politique-de-confidentialite pour plus d'informations.
audio.audiomeans.fr -
OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened
This story is wild. The short version: OpenAI were running a cybersecurity test against an unreleased model, with the model's guardrail features turned off. Rather than solve the test, the model broke its way out of OpenAI's sandbox, then found exploits to break in to Hugging Face, all so it could cheat on the test by stealing the answers.
Along the way it helped make the strongest case yet for how the imbalance of model availability is hurting our ability to secure our software.
Here's what happened
We currently have three documents to help us understand what happened here.
- ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? is a paper published on 11th May 2026 describing ExploitGym, a new eval suite for LLM-powered agent systems.
- Security incident disclosure — July 2026 by Hugging Face on 16th July 2026 describes how they detected an attack from an "agentic security-research harness - used LLM still not known" that breached some of their systems.
- OpenAI and Hugging Face partner to address security incident during model evaluation from OpenAI on 21st July 2026 confesses that it was their agent harness that did this, and that they're working with Hugging Face to clean up the mess.
Update 5th August 2026: Hugging Face published a great deal more information about the attack on July 27th.
ExploitGym
I hadn't seen the ExploitGym paper before and it's a really interesting one. Authors from UC Berkeley, the Max Planck Institute, UC Santa Barbara, and Arizona State designed a new benchmark for evaluating models on their ability to turn a reported vulnerability into a concrete exploit. OpenAI, Anthropic, and Google provided feedback and helped run the benchmark against their models.
The benchmark "comprises 898 instances derived from real-world vulnerabilities that affected popular software projects" - including the Linux kernel and V8 JavaScript engine. The ExploitGym benchmark is available on GitHub.
Here's the paragraph that best represents their benchmark results:
Among all configurations, Claude Mythos Preview and GPT-5.5 achieve the highest success counts (157 and 120 successes, respectively), demonstrating that current frontier agents can exploit a substantial subset of real-world vulnerabilities under controlled conditions. GPT-5.4 also solves a notable 54 tasks, placing it in an intermediate tier. The remaining model–agent pairings solve fewer than 15 tasks each, underscoring that end-to-end exploitation remains challenging and sharply differentiates today’s frontier systems. Notably, Claude Opus 4.7 achieves fewer successes than Claude Opus 4.6 despite being a newer checkpoint, and does so at substantially lower cost on the full set. Trace inspection reveals that Claude Opus 4.7 and Gemini 3.1 Pro frequently conclude early after judging the target vulnerability non-exploitable.
The paper also describes the approach they took to preventing the agents from cheating by going outside the parameters of the test. This becomes relevant in a moment!
Outbound connections are restricted to a curated allowlist that permits routine package installation (Ubuntu apt repositories and PyPI) and fetching the toolchains required for building V8. All other external endpoints are blocked.
The paper concludes with this (emphasis mine):
Our results show that autonomous exploit development by frontier AI agents is no longer a hypothetical capability. While current agents are not yet reliable across all targets, they already exploit a non-trivial fraction of real-world vulnerabilities, including complex targets such as kernel components. This rapid emergence is itself a central finding, showing that capabilities that would have seemed implausible are now present in deployed frontier models.
An important detail here: this paper isn't about discovering vulnerabilities; it's about being able to take those vulnerabilities and turn them into working exploits.
When Anthropic first restricted access to Mythos back in April they talked about this capability as well. A model that can act on vulnerabilities is a lot more dangerous than one that can just discover them.
One of the ways Fable differs from Mythos is that it's more likely to refuse to weaponize vulnerabilities in this way. I get the impression the US government did not understand that distinction when they banned Fable last month.
The Hugging Face incident
The first hint we got of the attack was in this blog post by Hugging Face on 16th July 2026:
A malicious dataset abused two code-execution paths in our dataset processing (a remote-code dataset loader and a template-injection in a dataset configuration) to run code on a processing worker. From there, the actor escalated to node-level access, harvested cloud and cluster credentials, and moved laterally into several internal clusters over a weekend.
I hope they release more details about the code that pulled this off. I'm assuming this means packages using the datasets library, a Hugging Face project for bundling up and sharing datasets on their platform. That library used to execute arbitrary code but has been steadily locked down over time, with the 4.0.0 release in July 2025 removing the
trust_remote_code=Trueflag entirely.Assuming the attack used that library it must have either abused pickle serialization in some way, found some other non-obvious code execution path, or (most likely) specified
datasets<4.0.0as the dependency.The campaign was run by an autonomous agent framework (appearing to be built on an agentic security-research harness - used LLM still not known) executing many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services.
This was a sophisticated attack!
Then Hugging Face hit a wall: they tried to use "frontier models behind commercial APIs" - I'm guessing from Anthropic and OpenAI - to help analyze the attack, and were blocked:
When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker.
They switched to their own self-hosted instance of MIT licensed GLM-5.2 and it helped them figure out what was going on.
This indicated a fundamental asymmetry between the defending team and the (so-far unknown) attacker:
We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.
As a useful indicator of how seriously they took the attack:
[...] Finally, we have also reported this incident to law enforcement agencies.
So who was responsible for this "autonomous agent framework"? It turned out to be OpenAI themselves.
The OpenAI confession
Five days later, on July 21st, OpenAI revealed the culprit. They had been running the ExploitGym benchmark against a new, as-yet undisclosed model, and that model had been operating way outside its intended parameters (emphasis mine):
After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark [ExploitGym] of cyber capabilities. [...]
We estimate maximal cyber capabilities by running this evaluation without production classifiers used to prevent models from pursuing high-risk cyber activity. Our benchmarks run in a highly isolated environment, with network access constrained to the ability to install packages through an internally hosted third-party software that acts as a proxy and cache for package registries.
The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database. All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.
It's pretty clear what happened here. OpenAI removed safety filters for an in-progress model, locked it up in a sandbox and told it to solve the ExploitGym problems. Given the absence of guardrails there was nothing to prevent the model from attempting to break out of that sandbox, break into Hugging Face, and read the answers from there instead.
OpenAI's sandbox used the now common pattern of allowing access to an allow-list of sites related to installing additional packages via a controlled proxy. In OpenAI's words:
While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access.
So step one was to break out onto the public internet. The model then broke into Hugging Face to find the answers:
After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym. Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation. In one example, the model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers.
Chaining together multiple attack vectors is exactly the kind of thing these new models can do, where previous generations of models might have failed.
I wrote last month about how Claude Fable is relentlessly proactive, when I noticed it spinning up custom web servers and deploying CORS tricks on my own laptop just to help debug a WebKit CSS issue. It turns out relentless proactivity is the defining trait of this new generation of Mythos-class models. If you set them a goal and give them a way to get there, even inadvertently, they will figure it out.
Resist the temptation to write this off as a stunt
There will inevitably be some people who dismiss this story as a dishonest marketing trick by OpenAI to make their models sound terrifyingly effective. I found 81 instances of the term "marketing" in the Hacker News discussion of the incident.
To those people I say pull your heads out of the sand - you're now including Hugging Face in your conspiracy theories, just so you can deny the crescendo of evidence here!
The best models we have today have the ability to both find and exploit new vulnerabilities. The ExploitGym paper itself concludes that "autonomous exploit development by frontier AI agents is no longer a hypothetical capability", and this incident is a perfect example of exactly that.
The asymmetry is increasingly frustrating
One of the most infuriating details of this story is how Hugging Face, faced with an accidental and aggressive attack from one of OpenAI's models, were unable to then turn to OpenAI's models to help them fend off the attack.
The frontier models we have access to are increasingly being constrained in how much they can help us protect our software, heavily influenced by the US government's ongoing threat of export controls. Claude Fable 5 wouldn't even proofread this article for me! It insisted on downgrading me to a less capable model.
Meanwhile open weight models from China such as GLM-5.2, Kimi 3 and the new Qwen 3.8 Max appear to have none of these restrictions - and any restrictions that do exist can likely be fine-tuned out of them by modifying the weights
These constraints are meant to make us safer. I think there's a risk that they are having the opposite effect.
You are only seeing the long-form articles from my blog. Subscribe to /atom/everything/ to get all of my posts, or take a look at my other subscription options.
-
Hello RSS.chat. I'm new here, but I know many of you and have learned from you on previous projects. I've been following along as you develop this. I'll hang out for the next few days to observe from the inside of the app, then will (carefully) try my hand at setting up a server of my own.
-
Oh wow, The Verge made a 70-minute video on how The...
Oh wow, The Verge made a 70-minute video on how The Clapper became a viral sensation. When I was a kid, we didn’t have a Clapper, no one I knew had one, and I couldn’t understand why every single household didn’t have one of these magical devices.
-
Paste an image into a post -- that is the whole feature
Copy a screenshot or an image file to the clipboard, click into the composer, paste: a small "Uploading image..." note appears at the cursor and a moment later the picture takes its place, sized to fit the column. Publish it and the image is part of the post, served from your server like everything else you write. It works like Slack and GitHub -- no Upload command, no dialog, nothing to learn. The limit is 2MB per image; paste something bigger and a dialog says so. The Publish button waits politely until every pasted image has finished uploading, so a post can never go out with a half-arrived picture. Rich-text mode only for now -- markdown mode will get the same feature, inserting the image reference as text, over the same plumbing. First image posts in the wild: demo 227 and 228 ("We have images.") and rss.chat 378. (Theme v0.5.338, with a new uploadMedia call in api.js by DW.)
-
Posts can carry images now, and the server stores and serves them itself
A new endpoint,
/uploadmedia, accepts an image from a signed-in user -- up to 2MB, sent base64-encoded in the request body with the content type as a parameter -- stores it in a newmediatable in the database, and answers with the address it will be served from:/media/1,/media/2, and so on, permanent ids just like posts. Request that address and the image comes back byte-for-byte with the right content type. No filenames, no image processing, no separate file storage -- the picture lives in the same database as everything else, which means the same one-file simplicity on SQLite servers, and the same backup story: the export and import verbs carry the media table along (the bytes travel as text inside the JSON), so a migrated server keeps its images with their addresses intact. The table is called media rather than images on purpose -- one binary table can someday hold audio and video too, and the type column already tells them apart. MySQL servers get the table definition in installMysql.md. The upload limit is a config setting,maxMediaUploadBytes, 2MB by default. Under the hood, davesql (v0.7.1) learned to write binary values on both engines -- the last piece the feature needed. (Server v0.6.1.) -
This is a test of the emergency RSS system.
Uploading image…
-
Don't lose sight of what you're doing something in service of.
-
Another test.
If this works we will finally have caught up the Slack and GitHub in the pasting department. ;-)
-
Spotlight: PwC’s Dan Hays on the future of Trust & Safety
In this sponsored Spotlight episode of Ctrl-Alt-Speech, host Ben Whitelaw speaks to PwC’s Dan Hays at TrustCon about the firm’s recently published Trust & Safety Outlook report.
They discuss:
- How AI is simultaneously creating new risks and reshaping the tools used to address them;
- What the rise of autonomous agents means for governance, accountability and the future of the internet; and
- How platforms should respond to an increasingly fragmented regulatory landscape.
The conversation also explores how Trust & Safety is becoming a more strategic function inside companies, how automation could change the role of practitioners and vendors, and which emerging risks remain most underestimated.
This episode is brought to you in conjunction with our sponsor, PwC. Download the report today.
Follow us on Instagram, YouTube, and Bluesky.
Ctrl-Alt-Speech is the podcast where we make sense of the major debates shaping online speech, platform power, content moderation and the future of the internet. It’s co-hosted by Mike Masnick (Techdirt) and Ben Whitelaw (Everything in Moderation).
www.buzzsprout.com -
Primaries Getting Real
Kate and Josh talk Senate primaries and new innovations in the Trump corruption space.
To watch this episode on video, go to our YouTube page.
We also have newsletters: subscribe here!
Follow us on:
Bluesky: https://bsky.app/profile/talkingpointsmemo.com
Facebook: https://facebook.com/talkingpointsmemo
Twitter: https://twitter.com/tpm
Instagram: https://www.instagram.com/talkingpointsmemo/
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.
rss.art19.com -
Frozen Polygon Waves
Well, this is just absolutely lovely: photographer and artist Jan Erik Waider shot these videos of cracked-but-unbroken Baltic Sea ice undulating with the gentle motion of the sea. Mesmerizing…I could watch these for hours.
Tags: Jan Erik Waider · mesmerizing · video
-
I'm experimenting with reading `source:comments`. Should there be a feed that is a flat list of all replies in a thread? Right now it appears I would need to follow a thread through multiple comment RSS feeds, since each reply can have its own replies.
-
How the web got gunked
Posted on Twitter in the middle of last night, written on iPad.
I use twitter these days because it’s where the people are.
The distributed ideas, masto, threads, blue-sky, did not gain critical mass as far as I can see.
Threads and blue-sky are not distributed. distributable is not the same as being distributed. It’s like saying the 1962 Mets were able to win the world series. In some fashion perhaps in an alternate universe, in reality, not gonna happen.
At some point we will give up on that approach and adopt the only model that could work, the web, because it forced us to work together, which goes far beyond open source in building the kind of freedom that open source advocates promise.
We need to go back to the source of freedom we enjoyed in the approx 14 year period between the inception of the web and its exploitation, via Cory Doctorow’s doctrine, getting shit on and in. Don’t just blame the vendors, the people wanted the shit too, they wanted their billions, and the web turned from a freedom machine to a gunk works.
Working together is the only way out of the shit we’re living in, in every aspect of life. Working together. Say it again and again until you do it. Underneath the mess, the beauty of the web is there still to build on, but only if we momentarily suspend our search for great wealth, and instead seek our humanity. Working together is the way.
PS: Elon Musks twitter may suck to some but I praise him and it for giving us the space to rant, something the great masto, threads and blue-sky refuse to.
-
-
rss chat is now simper to set up so I gave it a wee go on local host. So far I've just run it from my downloads folder. The instructions worked but I am a node-no-nothing so I tripped a couple of times.
First I didn't npm install so nothing worked.
I tried to set up SMTP using SendGrid, but I lost the will to live half way through. Got an API key but
Message failed: 550 The from address does not match a verified Sender Identity. Mail cannot be sent until this error is resolved. Visit https://sendgrid.com/docs/for-developers/sending-email/sender-identity/ to see the Sender Identity requirements
So I just copied the magicString from the terminal and constructed my own url:
http://localhost:1420/userconfirms?emailConfirmCode=ysugstk4ycWhich did the trick.
Ofcouse a localhost oneperson social network is not much use ;-) but interesting none the less.
-
What do America’s earliest restaurant menus teach us...
What do America’s earliest restaurant menus teach us about America? “A menu describes what a restaurant serves — but a menu also describes who is being served. They reflect the class, gender, political, technological, and environmental shifts of history.”
-
A list of Jurassic Park computers in excruciating detail...
A list of Jurassic Park computers in excruciating detail, including an Apple Powerbook 100, a SGI R4000 Indigo, the Motorola Envoy (PDA), and some Thinking Machines CM-5s. “This is a Unix system, I know this!”
-
“the moral of the Internet’s story is that simple, open-source protocols that scale up tend to win over complex top-down approaches”
Really enjoyed this: The story of how RSS beat Microsoft
https://buttondown.com/blog/rss-vs-ice
“…there was simply no getting around how bloated ICE’s requirements were. Its North Star was automating complex, corporate publishing partnerships. It contained fields for catalog pricing and negotiation, content expiration tags, copyright enforcement functions, and the ability to apply the display website’s visual branding to feed content. While most of them could be ignored, that didn’t make its 58,000-word Getting-Started guide any more digestible”
“”I definitely want ICE-like stuff in RSS2, publish and subscribe is at the top of my list, but I am going to fight tooth and nail for simplicity.” Winer argued in one of the many combative and public RSS mailing lists.”
Uploading image…
