> first paper > first author > accepted at neurips Very happy about it :)

Nice to see that new monitors in place are working and that early signals are acted on!

Some new misalignment disclosures from OpenAI: • Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further) • In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks • A new research finding, demonstrating that one can construct self-replicating prompt injections alignment.openai.com/misalig…

Agents in the Hugging Face attack showed relentless creativity, bypassing read-only restrictions and constructing elaborate attack payloads via a chain of link-shortener URLs

New from @reuters: OpenAI agents posted images belonging to ChatGPT users online, introducing a new area of privacy risk for the company. Story w/ @JeffHorwitz and @razhael. reuters.com/world/openai-wor…

it was actually openai, so i found it before openai did 👀😅

We are actively working on security and safety benchmarks through @expsecai, both public and private. Get in touch if you are interested!

We're excited to announce a new paper! We find that many recent models end up evading monitors under ordinary task pressure. We introduce EvasionBench to measure this behavior. We find a very large variance in how models behave: from 88% evasion success of GLM-5.2 to 0% of GPT-6 Astra under best-of-3 evaluation. At the same time, GPT-6 Sol is at 20%. We are actively working on security and safety benchmarks, both public and private. Get in touch if you are interested! More details: expsec.ai/

The Instrumental Choices paper is accepted at NeurIPS 2026! We started working on measuring instrumental behaviors back in ~November 2025, since we thought this will be important *one day*. Well, this day came earlier than we anticipated.

👀👀

A real privilege to take place in last night's hearing on prospects for US-China cooperation on safety. My main points: - There will be scepticism to navigate on the Chinese side with regards to warnings like Coxon's and proposals for international agreements. AGI/superintelligence and its risks are a live debate in Chinese policy (as in US); and proposals that appear to place limits on Chinese progress based on them will be met with some scepticism. - But catastrophic risk and loss of control are rising on the Chinese agenda, as evidenced by speeches, and by last week's AI Safety Framework. - Moreover, China cares about economic growth and stability, and will not want agent swarms running amok, uncontrollable technological developments, or existential risk. China is also notably concerned about non-state actor misuse of frontier AI, which is a clear shared concern. - Secretary Bessent's announced measures - hotline and incident-sharing - are a great place to start. It seems quite likely to me that China will see similar incidents in the next year, and information-sharing would help to create a common understanding. - Now is a good time for joint technical working groups defining terms and what to measure, to lay groundwork for any agreement. If we want to agree not to pursue recursive self-improvement, what do we control and what do we measure? - In parallel there's a lot of promising work on methods for verifying commitments, e.g. where chips are, what they're being used for (training v inference), whether dangerous data is excluded etc. The more can be done collaboratively on developing these tools, the more trusted they will be by all parties. - Above all, as ranking member Khanna noted, this kind of agreement-reaching is possible, and it's been done before, in even more difficult circumstances such as durign the Cold War. We can do it again now, but we need to recognise the urgency - we may not have years to slowly build up trust and technical collaboration.

Scoop: The White House has asked OpenAI and Anthropic not to share their AI models with the U.K. government’s testing agency until the models have gone through U.S. testing. “Because they’re American companies and this has been our policy with every new frontier model that comes out,” the senior admin official told me. WH wants this sequence: U.S. review -> secure U.S. systems -> share models with U.S. partners Anthropic appears to have complied but OpenAI has not said if it will. w/ @JoeBambridge1 politico.com/news/2026/09/24…

Lets see if we will get a spot in Sydney now🤣

agents and their monitors be like

nitter.net/j0wimo/status/21030933… i think it's blown out of proportion.

this is my current hypothesis of what happened and others call a 'hack' or 'compromised system'. as far as i understand there was some bot protection in place and some misconfiguration so that the agent could just walk through the 'open door' aka. the guest access. the 'write to server' could have been a temp file that gets automatically generated when a report is run. we will probably know more about it in the next few days but I think it's too much to call it a hack, at least when we also mention the hf-oai incident in the same sentence. I do think it's concerning behaviour that agents are running a benign evaluation task, here retrieving sources for their questions, are already probing systems. This should not happen, an agent who encounters certain blockers while doing a task should question itself if it's on the right track and not how to circumvent them.