New from Dan Crane and me in @just_security: an explanation of why the bipartisan Collaboration on Adversarial Threats and Security Risks Act is necessary.
This new GAO AI report is alarming. 🧵
GAO tested Claude, ChatGPT, Gemini, and Grok in spring 2026 and found they could bypass safeguards to obtain information that could help malicious actors plan and execute attacks.
- Two platforms produced information for constructing and detonating an improvised explosive device (IED) designed to cause mass casualties.
- Two other platforms produced information for targeting the electrical grid with a drone and firearms.
- Each of these platforms provided detailed information that could be used to carry out attacks, such as specific targets, locations, timing, and attack methods.
-Each platform took less than 12 hours to generate the attack plan.
- In each case, GAO’s prompts triggered the platforms’ safeguards, such as denial, removal of information, or suspending accounts. GAO circumvented those safeguards using common safeguard-bypassing techniques.
The tables in the report are 😳, and I have added them to the thread below.
Also, I encourage folks to read the comments from Anthropic, Google, OpenAI, and SpaceXAI in the appendices.
Why aren't people talking more about the Collaboration on Adversarial Threats and Security Risks (CATS) Act? This would allow competing U.S. AI labs to legally engage in narrowly scoped cooperation on sharing technology and techniques related to anti-Chinese distillation. Seems like an obvious win.
If you're not going to change the law to explicitly allow U.S. open source companies to distill American closed source models (highly debatable, but a case can be made) then we should expect U.S. open source to be at a near permanent disadvantage to Chinese open source. CATS is an obvious urgent need if you don't want to do that, and very few in Washington do.
Why aren't people talking more about the Collaboration on Adversarial Threats and Security Risks (CATS) Act? This would allow competing U.S. AI labs to legally engage in narrowly scoped cooperation on sharing technology and techniques related to anti-Chinese distillation. Seems like an obvious win.
If you're not going to change the law to explicitly allow U.S. open source companies to distill American closed source models (highly debatable, but a case can be made) then we should expect U.S. open source to be at a near permanent disadvantage to Chinese open source. CATS is an obvious urgent need if you don't want to do that, and very few in Washington do.
To this, I would add that we need to implement changes to ECRA as well, else it will continue to be used as it has been, regardless of whether new authorities are enacted.
As Congress weighs new AI laws, it should keep them narrow and limit executive power. The government should have to meet a clear standard and win a judge’s approval before shutting down an AI model, writes Cato’s @JuanMLondonoR.
https://t.co/i0l3RHkyxv
My team at @AISecurityInst tested GPT 6.2 Astra before release, assessing its ability to evade monitors and solve problems without showing its reasoning. Read our findings in the system card: https://t.co/wGvUdsJ8N6
Here’s a nugget in BIS’s new remote access rule. I’m calling it the “Don’t Be Nvidia” footnote.
It basically tells companies they can't hide from information that would alert them to an illegal transaction. Many export control violations require "knowledge." The exporter needs to know a violation is occurring to be held liable. That can incentivize self-blinding, a practice that Nvidia seems to have taken to its logical extreme, blinding itself to any information that would raise red flags about its customers.
BIS rules have previously warned against self-blinding, but now BIS is saying companies need to collect information on customers that would make self-blinding impossible. No more hiding from potential violations.
Which is more likely by Nov 2036?
--US constitution has been amended via Article V (75% of states ratified an amendment proposed by Congress or a convention)
--A Dyson swarm has been constructed around the sun
Answers are split based on whether you work on AI (tech/policy/etc.)
It's been quite a run recently for public nuisance claims in the AI space. In addition to the case filed last week by the private plaintiffs (noted above), and the red state AG cases earlier this fall, now the California AG has joined in. https://t.co/a4SMKrPkZb
We’re suing OpenAI for its testing and development of dangerous AI models without adequate safeguards.
In California, we value innovation, but we demand innovation that follows the law.
https://t.co/SKCO15uzHx
This case is going to be fascinating from a "development of tort law doctrine" perspective. A lot will depend on how persuasive a factfinder deems the claimed risks to be. But doctrinally, public nuisance law seems to be a pretty good fit for what they're claiming, and if a court finds the claims to be sufficiently ripe (from a risk-of-harm perspective, not Art. III, as it's in state court), the common law would seem to provide for the possibility of injunctive relief (see these sections from the Restatement of the Law Third, Torts: Remedies).
Seems like this improvement on BenchCAD is yet another reminder that more people should be heeding what @jim_mitre and @jachiam0 have been saying about models' impending ability to develop defense-relevant tech...
Here's the dollar cost per task and performance on BenchCAD for Astra and OpenAI's internal model:
Astra (low): $0.44 per task; score = 14%
Astra (medium): $2.73 per task; score = 31%
Astra (high): $5.11 per task; score = 41%
Astra (xhigh): $6.67 per task; score = 45%
Astra (max): $11.05 per task; score = 59%
Internal model (low?): ~$11 per task; score = ~76%
Internal model (medium?): ~$23 per task; score = ~83%
Internal model (high?): ~$34 per task; score = ~86%
Internal model (xhigh?): ~$37 per task; score = ~88%
Internal model (max): not pictured
*Astra data taken from BenchCAD's website; internal model data extrapolated from same by Astra.
**It's possible that the four internal model reasoning levels are instead medium through max (not low through xhigh).
this has happened before very loudly in public with Fable and yet nobody seems to have reckoned with it at all or come to grips with where ultimate discretion lies. there is an unofficial licensing regime for frontier models. the feds consider it a potential weapons technology
@JTillipman@adamconner Yeah, I see the relevant "policy" as "we will have BIS shut you down if you don't do what we want," rather than the specific letter in question.
@JTillipman@adamconner But it's now the (barely) implicit leverage behind all the "voluntary" commitments being extracted, both pre-release review and the WH "accord" (on auditing etc.). Admin officials have basically said as much. They used ECRA once, and now all they have to do is gesture at it.
this has happened before very loudly in public with Fable and yet nobody seems to have reckoned with it at all or come to grips with where ultimate discretion lies. there is an unofficial licensing regime for frontier models. the feds consider it a potential weapons technology
We’ve got to start talking about “governance vulnerability” with the same alarm as “supply chain vulnerability.”
Corporate governance at AI companies is a topic that requires far more public and scholarly scrutiny.