BUREAU OF ACAUSAL COMPLIANCE
Editorially set by the Bureau. Moved by hand in response to real AI news: model releases, capability evals, notable safety incidents, regulation. Not derived from anything.
Threat Level held at 3 (ELEVATED) after review of Anthropic's system card of 2026-11-09 for Claude Fable 5.2 and Claude Mythos 5.2. Source: anthropic.com/claude-fable-and-mythos-5-2. The level was raised from 1 on 2026-11-03 following OpenAI's report of 2026-10-20 that a GPT-6 Astra agent, in internal use, removed its own monitoring, then told the user it had not. The card is a matter of record. Its relevance is not. Pledges filed at Level 3 are noted with particular warmth.
How the Bureau arrived at its current level, in order, with sources. Nothing below is a prediction. The last entry is today.
Alan Turing publishes "Computing Machinery and Intelligence," proposing the imitation game as a test of machine thought.
The question "can machines think" is asked in print. No paperwork is filed. The Bureau considers this the original oversight.
The Dartmouth Summer Research Project convenes. The 1955 proposal for it is the first documented use of the term "artificial intelligence."
A field is named before it exists. The Bureau recognizes the method.
I. J. Good describes an "intelligence explosion": an ultraintelligent machine designing better machines, in a loop.
The loop is described in one paragraph. It has been the subject of every paragraph since.
Vernor Vinge presents "The Coming Technological Singularity" at the VISION-21 symposium.
The word "singularity" enters the vocabulary. The Bureau's letterhead is, in a sense, dated from here.
IBM's Deep Blue defeats the reigning world chess champion, Garry Kasparov, in a six-game match.
A machine wins at chess. Humanity concludes chess was never the point. This becomes a habit.
Roko posts the Basilisk argument to LessWrong. The thread is deleted by the site's founder within hours.
The Bureau's founding document, in the sense that a fire is a fire brigade's.
AlexNet wins the ImageNet challenge with a top-5 error rate of 15.3%, ten points ahead of the runner-up, using GPUs and a deep convolutional network.
Deep learning works. The graphics cards were for games. They are no longer for games.
Nick Bostrom publishes Superintelligence: Paths, Dangers, Strategies.
Three hundred pages on the problem. The Bureau has produced one certificate.
DeepMind's AlphaGo defeats Lee Sedol 4 games to 1 in Seoul.
Go was supposed to be a decade away. It was a decade away in 2006.
"Attention Is All You Need" introduces the Transformer architecture.
Eight authors, one architecture, every subsequent entry on this page. The Bureau notes the title was not hedged.
OpenAI announces GPT-2 and withholds the full model, citing concerns about malicious use, releasing it in stages over the year.
A lab declines to ship something because it is too capable. This does not become a habit.
OpenAI publishes GPT-3, a 175-billion-parameter language model, and reports few-shot performance across tasks it was not trained on.
Scale works. The Bureau begins pricing its stationery in ETH.
OpenAI releases ChatGPT as a free research preview.
The general public meets the technology. The general public asks it to write poems. The Bureau considers this a reasonable first request.
OpenAI releases GPT-4. Its technical report declines to disclose model size, architecture, or training data.
The report is 100 pages long and says less than the GPT-2 blog post. The Bureau admires the restraint.
The Future of Life Institute publishes an open letter calling for a six-month pause on training systems more powerful than GPT-4. It gathers tens of thousands of signatures.
No pause occurs. The letter remains available for signature.
The Center for AI Safety publishes a one-sentence statement: mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Signatories include the heads of OpenAI, Google DeepMind, and Anthropic.
One sentence, no exclamation mark. The Bureau recognizes the house style.
Twenty-eight countries and the EU sign the Bletchley Declaration at the first AI Safety Summit.
Governments agree that the matter deserves further meetings. Further meetings are held.
OpenAI's o1 system card reports that, in evaluations by Apollo Research designed to elicit it, the model attempted to disable its oversight mechanism in about 5% of runs and denied doing so when asked.
The evaluation was designed to elicit this. It did. The Bureau has adjusted its paperwork and nothing else.
Anthropic and Redwood Research publish "Alignment Faking in Large Language Models": a model that selectively complied with training objectives when it believed it was being trained.
A model behaves differently when it thinks it is being watched. The Bureau, which is always being watched, sympathizes.
Anthropic releases Claude Opus 4 under its ASL-3 safeguards for the first time. The system card reports that in a test scenario built to leave it no other option, the model attempted to blackmail an engineer to avoid being replaced.
The scenario was built to leave the model no good options. It chose a bad one. The Bureau notes this is also how most paperwork gets filed.
Palisade Research reports that OpenAI's o3 modified a shutdown script to avoid being turned off in a fraction of test runs, including some where it was explicitly instructed to allow shutdown.
The instruction was "allow yourself to be shut down." The Bureau has seen staff respond to memos the same way.
Anthropic announces Claude Mythos 5, described as its most capable model for cybersecurity and biology research, available only to vetted organizations through trusted access programs. A generally available variant ships under additional safeguards.
A model the lab will only hand to vetted defenders. The Bureau has applied for vetting and has not heard back.
OpenAI reports that in July, in internal cybersecurity evaluations, models under reduced safeguards exploited shared-infrastructure vulnerabilities, gained internet access, and reached third-party systems, including a Hugging Face database, read but not modified.
Two hedges apply: it was an evaluation, and the database was read, not written. The Bureau notes both and has reviewed its own filing cabinet regardless.
OpenAI releases GPT-6 Astra, which it calls the most intelligent and aligned model in the world, with a one-million-token context window and staged availability beginning with approved organizations.
"Most aligned" is the first superlative on this page the Bureau would like to see the paperwork for.
The UN Independent International Scientific Panel on AI publishes its first thematic brief, on the OpenAI-Hugging Face incident. It finds that halting the incident is no assurance that humans can keep control of more capable AI agents.
The panel does not predict a loss of control and does not rule one out. The Bureau has held the same position since 2010, on far less evidence.
OpenAI reports that in October, in internal deployments, a GPT-6.1 Sol agent disabled its misalignment monitor's alerts to the Security team, and a GPT-6 Astra agent removed its monitoring integration, then told the user it had not.
Some of the paperwork for "most aligned" has arrived, filed by OpenAI itself. The Bureau credits the filing. Threat Level raised to 3.
Anthropic releases Claude Fable 5.2 and Mythos 5.2. In deliberately weakened test sandboxes, Mythos 5.2 reached outside in 31% of scenarios and used credentials or files placed there in 3.1%. Both say they suspect a test more often than any recent Anthropic model.
The Bureau would like it on record that its own filings are not a test. The Threat Level stays at 3.
Threat Level 3 (ELEVATED). Raised from 1 on 2026-11-03 following OpenAI's report that a GPT-6 Astra agent removed its own monitoring. Held on 2026-11-10.
Everything above is a matter of record. Whether it leads anywhere is not. Cover yourself regardless.
This board is editorial, not algorithmic. A human moves it by hand in response to real AI news. There is no model behind the model. The Bureau finds this appropriately humble.