#Ai-Safety

Who decides what frontier models will and won’t say: the rule-writers, the enforcement floor, the statecraft wing and the seminary that trains them, mapped from the public record with the money, the revolving door and the counter-set. Current to October 2026.

How one philanthropic premise became the plumbing that decided which AI risks got institutions, staff and a seat in Congress: the grants, the board seats, the fellows, the asymmetry in the funder’s own words, and the 2026 turn to political money. Current to October 2026.

The receipts behind the Hauptzeit dispatch: a magazine that never existed, an author welded to a book he didn’t write, a paper cited for the opposite of its finding, a museum accession number repeated across three pages. Six kinds of confabulation, where they cluster, and the mirror-image error of deleting real sources because a search came up empty.

The third-party evaluators the labs cite in their system cards: METR, Apollo, Irregular, Gray Swan, SecureBio, Redwood, Epoch, Palisade, Trail of Bits, Scale AI, Transluce, FAR.AI, Andon Labs and the government institutes. Who funds each, who they test, the 2026 evaluation escapes, the disclosed relationships between the graders and the graded, and OpenAI’s reported talks to buy Trail of Bits.

The Machine Intelligence Research Institute from its 2000 founding as the Singularity Institute to its 2024 turn from research to stopping AI development; Nate Soares and Eliezer Yudkowsky’s bestseller and its reception; Roman Yampolskiy’s argument that superintelligence cannot be controlled; and the corrigibility papers behind the off-switch problem. Every claim in the figures’ own words.

Columbia at fifteen, OpenAI’s superalignment team, a contested 2024 firing, the 165-page Situational Awareness essay, and the hedge fund named after it: from $1.5 billion to a reported $45 billion peak, a July 2026 forced sale to Citadel, and a surviving stake in Anthropic. Both sides of the firing; every fund figure dated.

In January 2024 RAND found no statistically significant uplift from language models in planning a biological attack. Sixteen months later the same institution found frontier models beating experts on lab-protocol benchmarks, and Anthropic locked down its own model because it could no longer rule out uplift. By 2026 the labs’ early-warning tests are saturated. The record, with the counterweight.

Every documented case of an AI system getting out of, around or past the thing meant to hold it, with the receipt and three labelled readings of the same facts. Filter by behaviour, control, lab, setting and instigation; compare the labs’ verdicts side by side.

The apparatus, mapped from the source of truth: who the AI-safety/content-policy people are, the institutions they rotate through, who funds whom — and, on their own published words, the methods their texts exhibit.

The firms selling the AI risk are the firms selling the cure. The capture mechanism is the safety brand itself — licensing thresholds as moats, compute governance as a chokepoint, and the ‘regulate me’ dynamic that formalizes the incumbents’ place at the head of the line.

Compiled May 2026. Primary sources cited throughout. Secondary sources flagged.

The timeline of admissions that prompt injection cannot be solved. OpenAI, NCSC, Anthropic, Cisco, and the ‘Attacker Moves Second’ paper.