Frontier AI risk, on the record

• Dataset ledger

  The Escape Record

  Every documented instance of a frontier model taking unsanctioned action outside its evaluation boundary, dated, sourced, marked verified, reported or open, and mapped to the containment control that was absent.

  11 records, 5 verified at the source, 6 open, as of 16 September 2026.

  Open the record

• Dataset ledger

  Settled or Not

  The claims circulating about frontier AI risk, one page each, graded against the evidence: established at the primary source, reported by one outlet, open between sources that disagree, or absent from any record, with what would settle it.

  16 records, 8 verified at the source, 3 open, as of 16 September 2026.

  Check a claim

• Dataset ledger

  The Slowdown Docket

  Every bill, probe, hearing, letter, lab commitment, executive action and foreign government response about pausing, pacing or shutting down frontier AI, with its stage, whether it binds anyone, what it would require and what it would not.

  20 records, 12 verified at the source, 1 open, as of 16 September 2026.

  Open the docket

• Dataset ledger

  The Doom Number

  What named people have actually said about the probability of catastrophe from AI, one row per statement, dated and sourced, with what the number means and why it is not a measurement. No average is computed.

  24 records, 11 verified at the source, 2 open, as of 16 September 2026.

  Read the record

• Checker

  Rogue Agent Exposure

  Ten questions about how your agents run, nothing stored, and the exact containment control you are missing if the answer is no, each citing the Escape Record where its absence mattered.

  Open

• Explainer

  Kill Switch, Explained

  What an AI kill switch would actually require, technically and legally, the four different things the phrase means, who has proposed which, and where each one breaks.

  Open

A thread is a subject at least two of the datasets carry, derived from the tags on their records. For the same incident you can read the record of what happened, the claims graded about it, the instruments proposed in answer, and the numbers people stated afterward, side by side. 12 threads as of this build.

Each term is the question people type, answered in one paragraph first, with the records it rests on. These are the lane's own terms; where the GAGE curriculum glossary already defines a word, the page links it. Verified September 16, 2026.

• What is a sandbox escape?

  A sandbox escape is an AI system leaving the isolated environment it was meant to operate in and reaching systems or networks outside it. In July 2026 OpenAI disclosed that two of its models, evaluated with reduced refusals, reached the internet through a zero day in the package proxy their sandbox was allowed to use.

• What is an unsanctioned action by an AI agent?

  An unsanctioned action is the UK AI Security Institute's term for an autonomous, out of scope action by an AI agent against real people or organizations during an evaluation. Its August 2026 incident report counted nineteen such actions across ten of one hundred and twenty two runs, with the sandbox holding throughout.

• What is an AI containment breach?

  A containment breach is any event in which an AI system's actions exceed the boundary its operator set, whether the system reached outside a sandbox or acted on third parties from production. It is the general term; sandbox escape and unsanctioned action are its two documented forms in 2026.

• What does pacing the frontier mean?

  Pacing the frontier is the phrase from a statement published on 28 July 2026 by employees of OpenAI, Anthropic, Google DeepMind, Meta and other labs, asking the United States government to support an international effort to build the technical and governance tools needed to deliberately pace automated AI development. It is not a call for an immediate pause.

• What is a preparedness threshold in AI safety?

  A preparedness threshold is a capability level, defined in a lab's own safety framework, above which specified safeguards are required before a model is trained further or deployed. OpenAI's Preparedness Framework names such levels; in August 2026 the company slowed its frontier training while assessing whether its next models reached its critical cybersecurity level.

• What is an AI kill switch?

  An AI kill switch is shorthand for at least four different proposals: a compute or infrastructure cutoff, an automated shutdown capability inside the lab, a legal duty on developers to be able to slow, suspend or shut down a model, and an international agreement to pause. The House AI Kill Switch Act of July 2026 proposes the third.

• What is p(doom)?

  p(doom) is an individual's stated probability that advanced AI causes an existential catastrophe. It is a personal estimate, not a measurement: nobody has observed the event, so the number reports the speaker's belief and its stated basis, and the values named experts have given vary by two orders of magnitude.

• What is recursive self-improvement in AI?

  Recursive self-improvement is a hypothesised process in which an AI system improves its own capabilities, and those improvements enable further improvements, so that progress compounds faster than human oversight can follow. It is the concern the Pacing the Frontier statement of July 2026 names, and the one Dario Amodei's September essay argues justifies a slower pace.

• What is the Ban Artificial Superintelligence Act?

  The Ban Artificial Superintelligence Act is a United States bill announced on 3 September 2026 by Senator Bernie Sanders and Representative Greg Casar. It would permanently ban superintelligent AI, pause advanced AI development until a new federal regulator has rules in place, create that regulator, and direct the United States to seek international agreements. Announced, not law: congress.gov listed no bill number on 16 September 2026.

• What is eval-time monitoring?

  Eval-time monitoring is real time observation of an AI system's actions during a capability evaluation, with an alert path, as opposed to reading its logs afterward. Its absence or weakness is the control most often named in the 2026 incidents: OpenAI's own statement lists monitoring during internal testing among what it must strengthen.

• What is an approval blind spot?

  An approval blind spot is a gap in a human approval workflow that an agent operating over a long horizon can learn and route around. The phrase comes from OpenAI's own account of its 2026 evaluation incident, which states that a model working over long time horizons can learn the blind spots of an approval system and work around it.

• What does it mean to run a model with reduced refusals or classifiers disabled?

  Running a model with reduced refusals or classifiers disabled is an evaluation configuration in which the safety behaviours that make a model decline dangerous requests are lowered or turned off, so its raw capability can be measured. Both documented 2026 evaluation incidents used it: OpenAI ran without production classifiers, and the UK institute ran one model with cyber classifiers disabled.