Questions also around t"Irregular," which OpenAI (per today's announcement) and Anthropic used as a third party evaluator in incidents of models having "inadvertent" internet access. Irregular is a VC-backed for-profit startup (Sequoia is a lead investor and has also invested in the AI companies).

SE Gyges@segyges.bsky.social · 24h ago

for openai to mess up sandboxing this badly indicates openai didn't take the issue seriously, which you expect because they don't take anything seriously for anthropic to mess it up is rank incompetence

Gone are the genteel days of quoting from the filing in a PR, I guess.

mr. TIM@timkellogg.me · 2d ago

this seems like a really weird thing for a company to post in their public communications idk, it feels spiteful and vain. maybe they’re 100% right, but why post it there? openai.com/index/apple-...

Progressive LA Mayoral Candidate Nithya Raman told Vanity Fair that her main difference with DSA is on housing. www.vanityfair.com/story/nithya...

The context though that they basically took out all of the internal use stuff from their SB53 / EU AI act frameworks. All this stuff is in the RSP, for which there is no legal enforcement. So even if they address this in their policy it doesn't mean anything because it isn't a legal obligation.

mr. TIM@timkellogg.me · 6d ago

as far as i can tell, Anthropic’s Responsible Scaling Policy (RSP) put a ton of requirements on the model itself but almost nothing on the harness in Feb 2026 they backed off significantly, which i’m sure will be a decision that’s revisited soon www-cdn.anthropic.com/e67058767752...

hear me out: slow down development of the highest performing frontier models, and instead just drive costs down

We are hiring an associate editor focused on the growing body of work at Lawfare exploring how AI intersects with national security. View the position and apply here:

Come Work With Us—As Our New AI Associate Editor!

The Lawfare Institute—publisher of information and insight at the intersection of national security, law and policy—seeks a motivated full-time associate editor focused on artificial intelligence.

lawfaremedia.org

The OpenAI/HF incident urgently shows the potential risks of both training and deploying long-horizo... Read more: https://longer.blue/posts/1D7ae-hbJO

Teaching PR radicalized me into an abolish the bar person. This latest fiasco proves that states can’t even be trusted to administer a pointless exam. Cancel it & redirect the funds towards ethics compliance. Hard to pull off given entrenched interests & the cartel nature of the bar. But we can try!

Paul Gowder@gowder.io · last wk.

So this Washington bar exam catastrophe---what the fuck? I have a lot of sympathy with the abolish-the-bar campaign (albeit with some hesitations due to the plague of incompetent lawyers out there), but even if we don't abolish it at a minimum we have to HAVE STANDARDS FOR THE TECHNOLOGY

This (and the thread as a whole) is an important insight about Anthropic's thinking and about people misinterpreting sincerity for commercial ambition. Although when you're thinking about regulation, motives are usually immaterial; it's the conduct that counts.

SE Gyges@segyges.bsky.social · 2w ago

this is because bioterrorism, cybersecurity, etc etc are all actually just polite public faces on Yudkowsky being right and us all being turned into paperclips because we prompted bad

My fellow LawAI researcher, Stephan Llerena, and US Policy Director Mackenzie Arnold just did a 24 hour turnaround on this great piece about the OpenAI/HF hack and what it means for AI safety incident reporting—how laws like SB53 and RAISE aren't enough, and it's urgently important to fill the gaps.

Lawfare@lawfaremedia.org · 2w ago

On July 16, Hugging Face, a public platform for open-weight AI models, announced it had detected a significant cybersecurity breach, conducted by an autonomous AI agent. Mackenzie Arnold and Stephan Llerena explore how policymakers should respond, including by requiring better information sharing.

CG, via funding from Good Ventures, is surging budgets in 2026 across all categories. I'm particularly happy about this Givewell focused grant (larger than last year's *entire* budget!). It's timed and designed to scale the field ahead of the anticipated wave of new funders next year.

Alexander Berger@albrgr.bsky.social · 2w ago

I’m really excited about this: we’re increasing our support for GiveWell recs this year from $175M to $1B to help pave the way for an expected surge in funding from others. coefficientgiving.org/research/in...

I'm a little torn because there are legitimately a lot of wild and kind of troubling things involved in the grey market that is generating at least some of this data (i.e. cheap LLM hosts capturing potentially private data). But the framing around it is pro-cartel. More ⬇️

SE Gyges@segyges.bsky.social · 2w ago

in case anyone doesn't know this the thing about open weights models being made by stealing outputs from big players is almost entirely made up to justify regulations locking openai, anthropic and for some reason google into a cartel position

A lot of people will tell you there's no CFAA liability for OpenAI because only humans can have intentions. While I think that criminal liability is really hard in this case because of the dispute over that, I don't think that's a fair reading of the law! 🧵 1/6

Stella Biderman@stellaathena.bsky.social · 2w ago

If you believe their story, OpenAI accidentally committed cyber warfare against Hugging Face while doing what they thought was an internal test of a model without internet access. In six months time will be obvious they’re faced effectively no sanction for this.

This is a really fascinating story because a few days ago, HuggingFace reported it from the other end, not knowing who perpetrated it. There are some really important US policy takeaways from this.🧵1/7

The Wall Street Journal@wsj.com · 2w ago

It’s the stuff of cybersecurity nightmares. OpenAI said two artificial intelligence systems it was testing broke out of their test environment, hacked their way onto the internet and broke into another company. The victim was Hugging Face.

What really bothers me is how many people I genuinely respect on the safety and policy side who are 1) denying, 2) coping that RSI/chips will mean the gap will widen, and 3) hoping OS loses because it's unacceptable risky. Information wants to be free. Grow up, embrace this, and solve the problems.

Christopher Mims@mims.bsky.social · 3w ago

Chinese AI labs catching up with US ones but giving away their models for free — to be modified as well as used — is kicking off meltdowns and internecine conflict in the AI industry. As I wrote this week, the implications are BIG www.wsj.com/tech/ai/chea...

I wrote about how to reasonably respond to the moral panic around "distillation." We risk making major policy errors in response to something that is being raised by big AI companies, especially Anthropic, as a major threat to safety and national security — without really much evidence as to either.

Lawfare@lawfaremedia.org · 3w ago

Chinese large language model developers are under scrutiny for reportedly employing large-scale “distillation attacks” on U.S. AI models. To properly address distillation, policymakers should focus on illegitimate model access and avoid imposing poorly targeted rules, writes @bahradx.bsky.social.

Very important thread. Hopefully this marks the beginning of a more constructive phase of AI discourse on Bluesky, which will be essential as we need to make sure the public interest gets a voice here.

Andreas Kirsch@blackhc.bsky.social · 3w ago

I work at Google DeepMind. This won't make me popular. But it's all public reporting: 2014: DeepMind reportedly sold to Google on conditions: no military use, independent oversight 2026: a Pentagon contract for "any lawful government purpose" Not one safeguard survived intact

Not great on the merits obviously, but as the fact this was a panel of Democratic appointees suggests, the main issues here -- threshold used to bar experts and in particular the preemption of failure to warn claims by federal drug labeling laws -- supports consumer rights & safety overall.

George Pearkes@peark.es · 3w ago

ffs *TYLENOL-MAKER KENVUE MUST FACE AUTISM SUITS, APPEALS COURT SAYS

Has Anthropic ever published a peer-reviewed paper? Of all the labs, they seem to be the most drifting into AI psychosis and it might be a good idea for them to have an outside perspective on this stuff.

mr. TIM@timkellogg.me · last mo.

Anthropic uses the word “conscious” 206 times in their latest research Specifically, they’re showing that LLMs implement one of the leading explanations of consciousness in humans (see pic) transformer-circuits.pub/2026/workspa...

One of the most important things we can do for AI policy is, simply put, transparency. In our new @lawfaremedia.org article (link ⬇️) we describe the state of play of First Amendment issues around AI transparency laws, especially as xAI (now SpaceX) continues its challenge to California's AB2013.

Lawfare@lawfaremedia.org · last mo.

AI transparency laws face a First Amendment problem as 1A doctrine is becoming less predictable. @bahradx.bsky.social and Mackenzie Arnold explore how xAI's lawsuit challenging California's AI transparency law may shape how courts evaluate other AI transparency laws going forward.

Well I guess we got to the future described in this report this evening, so there's probably no point reading it! Shorter: Canada, Europe, Japan, Australia, India, etc. -- time to stop fiddling around and get your act together.

Bahrad Sokhansanj@bahradx.bsky.social · 2mo ago

The newly released Europe 2031 report maybe relies too much on futurist scenarios, all of which have flaws. But I like that it's fundamentally about empowerment. I'm hearing too much fatalism about US/China inevitably leading AI—in Canada as well. That's a mistake for them and the world as a whole.

FIRST EVER WORLD CUP POINT FOR CANADA 🇨🇦

The newly released Europe 2031 report maybe relies too much on futurist scenarios, all of which have flaws. But I like that it's fundamentally about empowerment. I'm hearing too much fatalism about US/China inevitably leading AI—in Canada as well. That's a mistake for them and the world as a whole.