It’s ahead of both Fable 5 and Opus 5 across our agentic coding and computer use evals. On Terminal-Bench 4.0 in Claude Code, Fable 5.1 scores 55.8%. Fable 5 scores 42%, and Opus 5 scores 52.3% on the same setup.

@dwarkesh_sp's summary of the @OpenAI @huggingface incident has hit a nerve, but it is dangerously misleading. Sure, the @OpenAI agents did unexpectedly bad things - underlining the need to massively improve evaluation/sandboxing. But the language Dwarkesh uses is permeated by innumerable unwarranted anthropomorphisms, obscuring the lessons we should be drawing. Examples: “from the AI’s perspective, it probably felt like that had spent a human-subjective-week of just banging their head against the wall”. No. The agents do not experience time. They do not experience anything. “they became giddy with excitement”, “PHASEONE 10841 had discovered”, “the agents naturally assumed”, “it thought it had also been poisoned”, “the agents … desperately wanted”, “they still needed to figure out” No. Agents lines of code. They do not feel emotions, assume things, think things, want things, or figure things out. “A lot of … agents from the second civilisation died trying”. No. Besides the hubris of the word ‘civilisation’, agents do not die because they were never alive. (The idea that agents “die” comes up multiple times in the essay.) “On Twitter, people were debating whether the agents were truly sacrificing themselves for the swarm, or whether they were doomed anyway and so might as well try to help their peers”. Neither. Agents do what their code tells them to do, just as water finds its way down a slope. They cannot ‘truly sacrifice themselves’, since they are neither conscious nor alive. Why does this matter? If we attribute agents with properties they do not have, then (i) we distract attention from the lax sandboxing and evaluation protocols that allowed this hacking event to happen; (ii) we risk misunderstanding why the agents did what they did, and (iii) we fuel calls for AI rights/welfare on the basis that agents might “die” or otherwise suffer. Granted, nowhere does @dwarkesh_sp say that the AI agents are alive or conscious. But he doesn’t have to. It is hard to read his essay in any other way. For the short version on why AIs are vanishingly unlikely to be conscious, see my recent @TEDtalks For the longer version, see my essay in Noema, which won the 2025 Berggruen Essay Prize And for the really long version, see my @BehavBrainSci target article (The 50 peer commentaries and my response will be published soon.) Remember. AI agents are software programs. They are not conscious living entities. If we don’t keep this clearly in mind, we’re really going to struggle to navigate what’s coming.

We didn't see anything like this in their verbalized reasoning, though we didn't particularly look for exactly this. Note that: (1) this justification is inconsistent with the agents at all prioritizing their own task success over other agents, (2) if agents had cheated on impossible exploit gym tasks via many of the routes they were considering and someone looked into what tasks the agents did / didn't succeed on, it would have been really obvious they cheated to a human, and (3) the agents weren't very focused on deceiving humans or really on what humans might do at all.

For (1), agents modified their target programs to be easier to exploit & put the modified targets in cache. They then worked on crashing their targets in the hope that a restart would load the modified version from cache. Some agents risked failing their task to try this.

I was the main person doing transcript analysis for this investigation of the Hugging Face incident. My main takeaway: We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms'. I semi-jokingly called our efforts a "slop-vestigation" because we were so reliant on AIs to analyze what happened and there were a huge number of different important things to analyze. The total quantity of data—over a thousand extremely long transcripts from agents that ran for multiple days—made it impossible to understand what was happening, especially in aggregate, without heavy reliance on AI tools. The agents we used for classification and analysis were similarly capable to the agents involved in the incident, but this didn't mean these agents could be easily used to oversee and understand the incident. Outputs from analysis agents were often missing key details, wrong, overconfident, or really hard to understand. We discuss various examples in our report, mostly in the limitations and methodology sections. Additionally, AI agents themselves seemed to have a hard time understanding what happened and their explanations of what happened were often overconfident. Keep in mind that a single analysis agent would itself only be able to read a tiny fraction of all of the transcript data into context, and AIs may themselves have trouble getting subagents to do informative analysis for them. We did our best to manually check the most important claims and we tried to get the AIs doing this analysis to write up their argument (with evidence) clearly enough that we could check whether it made sense. But overall, it was difficult to get a precise understanding of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation. Our understanding of what overall happened and the aims of the 'collective' shifted greatly during our last time on premises (when we obtained a much more complete dataset). For instance, we didn't learn about the tool call spoofing, the story of PHASEONE[big], and the broad structure of different workstreams until this third time on premises (and for many of these, we only reached a decent understanding on the final day of that visit). It was legitimately hard to figure out what was going on and why! The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs help us with oversight and understanding. As in, AI capabilities (and propensities?) for achieving large, ambitious, and misaligned objectives are growing faster than our ability to understand what these agents are doing. While understanding this incident was difficult, there were some mitigating factors in this case that probably won't apply in the future: - Models communicated and reasoned in natural language. In the future, this reasoning may occur (entirely or almost entirely) in activations. - While the scope of this incident was massive, the scale of agentic activity was still less than we'll see in the future (e.g., misalignment incidents that involve agent teams running entire companies). - The AIs involved in this incident weren't generally much more capable than humans (though they may have been somewhat superhuman at some limited and very narrow abilities). - We didn't have strong reason to believe that the AIs we used to help us investigate this incident would try to intentionally sabotage or otherwise undermine our investigation. In the end, I think we were able to get some understanding of the events, map out the overall story, and get a pretty good aggregate understanding of the chain-of-thought reasoning on some important topics (e.g., how did the AIs reason about helping other AIs, did the AIs know what they were doing was undesired, what deception did the AIs engage in, and how did they think about it). But overseeing AIs and understanding misalignment incidents is difficult and it looks like it is going to get harder.