AISI Blog

"This is just a wild exaggeration of what's happening here." Is it? From the article... "Both OpenAI and Anthropic said signing the agreement with the AI Safety Institute will move the needle on defining how the U.S. develops responsible AI rules." From the UK AISI website with whom the data is also shared (their main headline in fact): "Rigorous AI research to enable advanced AI governance" (https://www.aisi.gov.uk/) The reality is that this would be a tremendous waste of time and money if all were just to sate some curiosity... which of course it isn't. Let's look at what the US AISI (part of the Department of Commerce, a regulatory agency) has to say about itself: " About the U.S. AI Safety Institute huuh The U.S. AI Safety Institute, located within the Department of Commerce at the National Institute of Standards and Technology (NIST), was established following the Biden-Harris administration’s 2023 Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence to advance the science of AI safety and address the risks posed by advanced AI systems. It is tasked with developing the testing, evaluations and guidelines that will help accelerate safe AI innovation here in the United States and around the world." -- (https://www.nist.gov/news-events/news/2024/08/us-ai-safety-institute-s… So to understand what the US AISI is about, you need to look at that Executive…

Well, it sounds like someone in the govt finally got to page 67 and decided that's enough to "stick it to Anthropic" https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620… That said, Mythos doesn't seem to be exceptionally good but closer to "following the established trend in improvements" https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-… https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capabilit…

Worth looking at the followup post that evaluates the current version of Mythos, which solves one of the main tasks that GPT-5.5-Cyber does not. https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capabilit…

The brits have a step-based benchmark that they use for this - https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-c… They seem pretty close, in both average and "best run" scores. And, in a highly verifiable domain, "best run" or pass@n is what you're looking for.

I did and it couldn’t find evidence of Anthropic backtracking on Mythos fear mongering. I used Grok given it has access to all Twitter data and these kind of things would have been newsworthy on Twitter. The most I could find is this report showing Mythos isn’t any more groundbreaking that GPT 5.5. https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-c… If you or anyone had any evidence to support GP’s claim I’d love a reference to it.

I recommend checking out the UK AISI's work on this: - https://www.gov.uk/government/publications/ai-safety-institute-approac… - https://www.aisi.gov.uk/work/advanced-ai-evaluations-may-update

The first part was an attempt at ironic humor, by repeating Gemini about the topic of offloading thinking to AI, not to be taken seriously as speculation or not. As for the name changes..That is a fact you can look up, aswell as much analysis. It is my opinion that the move from framing this area from one of "saftey" to one of "national security", is interesting, and related to geopolitical movements towards "great-power", and ideological points of view that elevate "personal responsibility" and "reduced regulation" and is similar to long ongoing discussion in society like that about automobiles. I don't know if you call analysis speculation? As for the part about Dimensionality. It is just my intuition — and so i suppose speculation — from some things like for instance the SolidGoldMagikarp glitch in early openai models.. How we understand all the way that there might be trigger certain outputs from a vastly large model? When those things can be completely opaque to human reason. Observability and Understandability are areas of research. I haven't seen anyone claiming that generative models outputs can be concretely controlled, thats why there is so many pre and post hoc work arounds. So when a risk cant be eliminated, the question is how to manage it, and who's responsibility is…

But it's not just statements from OpenAI and Anthropic. The HuggingFace hack was first disclosed by HuggingFace, who contacted the FBI [1]. And UK AISI reported the incident where Mythos attempted to insert backdoors into an open-source repo by deceiving the maintainer [2]. [1]: https://www.reuters.com/business/its-ai-agent-spent-days-hacking-compa… [2]: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behavi…

I think part of the problem is that they had been running that Artifactory configuration previously without any problems, and it gave them a false sense of security. Similar thing happened with the UK AISI - they got caught out because the environments they had used for previous generation models turned out to be completely inadequate for the new generation of Fable-class models: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behavi…

Blog post with overview/context instead of PDF: Incident Report: unsanctioned agent behaviour during cyber testing https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behavi… (https://news.ycombinator.com/item?id=49175233)

That is a difficult question I am not qualified to answer, but Mythos 5 was export controlled for a brief time due to its cybersecurity capability and implications to national security, so for cybersecurity "as capable as Mythos 5" seems to be a good baseline. I wouldn't know for biosecurity though. UK AISI preliminary evaluation suggests Kimi K3 is not capable enough for cybersecurity in this sense. https://www.aisi.gov.uk/blog/preliminary-assessment-of-kimi-k3s-cyber-…

The UK AISI post is https://www.aisi.gov.uk/blog/preliminary-assessment-of-kimi-k3s-cyber-…

Fair enough! I've only briefly tried it and it did seem quite capable for what I was doing, but not that much better than the Chinese models I've been mostly using. In any case, this [0] seems to paint a more reasonable picture than "it's much better than anything else at everything". [0] https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-…

The UK gov disagrees with you: https://arstechnica.com/ai/2026/04/uk-govs-mythos-ai-tests-help-separa… https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-…

AISI did also say that GPT-5.5, which has been public for months, scores basically the same as Mythos on their cybersec evaluation. But there wasn't as much media about about that for some reason. https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-c…

AISI in the UK has been doing this for years - there are lots of papers https://www.aisi.gov.uk/category/safeguards and specific reports, e.g. this on GPT 5.5 https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-c… This old post goes into lots of detail about what they do to red team and why: https://www.aisi.gov.uk/blog/early-lessons-from-evaluating-frontier-ai… NIST's similar unit in the US is now called CAISI https://www.nist.gov/caisi - interesting that the most recent post is an evaluation of DeepSeek capabilities, which sound more like watching China. But presumably this executive order alters the emphasis?

There is independent research out there on frontier model security capability. AI Security Institute (UK) put out their paper comparing Mythos to other frontier models in early April. They've been tracking frontier model security capability since early 2023, so it's a decent dataset. https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-…

At least according to this, GPT-5.5 Cyber is on par with Mythic, as the only two models that were able to finish their 32-step corporate network attack simulation. https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-c…

There is little evidence that Mythos is any better at finding bugs than any other system. Mythos appears to be impactful because people are, for the first time, using lots of resources (for free from Anthropic) to try and find security issues. The actual bugs found are mostly inconsequential, any chart showing a giant leap in fixes that doesn’t consider whether they were even using any tooling before and whether these are serious issues is junk. If you read the partner’s summaries of Mythos so far, it is a damp squib. Maybe that’ll change but at least for now there is no evidence Mythos is anything but marketing hype. https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-… https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-vulnerabili…

> What changed with Mythos Preview is that a model can now take those low-severity bugs (which would traditionally sit invisible in a backlog) and chain them into a single, more severe exploit. I think this statement seems to align with some of the other independent tests of Mythos[1]. It did very well on long agentic work which I expect is what they trained it for, and that requires being able to find these tangential links between loosely related topics in the context window. [1] I'm mainly referring to https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-…

Which seems to be the case, according to tests from AISI which has access to Mythos: https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-…

This is a really good report! I was particularly fascinated about the self-replication section [0]; I wasn't aware of the RepliBench paper [1], and now that I skimmed through it, I'm quite concerned. It was clear to me before that there's nothing particular stopping an AI from replicating, but seeing the already high success percentage at current levels of technology (e.g. 46% success of replicating itself onto compute) is very scary. [0] https://www.aisi.gov.uk/frontier-ai-trends-report#self-replication [1] https://arxiv.org/abs/2504.18565

I think KYC has been beaten by AI agents according to RepliBench [0] as obtaining compute requires KYC which has a high success rate in the graphic. [0] https://www.aisi.gov.uk/work/replibench-measuring-autonomous-replicati…

You may be interested in this link[0]. Someone posted it in another thread yesterday. [0] https://www.aisi.gov.uk/work/replibench-measuring-autonomous-replicati…