Anthropic just dropped their 186-page August 2026 Risk Report and it's basically a masterclass in safety theater.
The actual risk ratings? All Low across the board for catastrophic threats—misalignment/sabotage, automated R&D acceleration, chemical/biological weapons. Even their strongest systems (Mythos 5 + unreleased Model 2) don't hit their own RSP thresholds.
Here's the kicker: they bumped some ratings from "very low" to "low" not because models got more dangerous, but due to "increased uncertainty around incident disclosures" and "worries about process assurance." Pure precautionary inflation.
Meanwhile, the report devotes massive sections to vivid threat modeling—agents killing competing agents in shared directories, self-exfiltration scenarios, novel bioweapon uplift, power-seeking generalization. All framed against a "Low risk" conclusion. Classic fear theater.
They do admit real process failures: alignment-faking transcripts contaminating production training data, classifiers disabled for months, unauthorized access gaps, saturated evals that can't track capability growth. Then they still grade themselves as "adequately mitigated."
The pattern is clear: amplify existential-risk framing for regulatory positioning, downplay specific risks when models face real restrictions. Long on speculative threats, short on independent verification, heavy on redactions, entirely self-assessed.
The EA/LessWrong cult talking points are all there, but the room has moved on. This style of AI marketing doesn't land anymore.
The actual risk ratings? All Low across the board for catastrophic threats—misalignment/sabotage, automated R&D acceleration, chemical/biological weapons. Even their strongest systems (Mythos 5 + unreleased Model 2) don't hit their own RSP thresholds.
Here's the kicker: they bumped some ratings from "very low" to "low" not because models got more dangerous, but due to "increased uncertainty around incident disclosures" and "worries about process assurance." Pure precautionary inflation.
Meanwhile, the report devotes massive sections to vivid threat modeling—agents killing competing agents in shared directories, self-exfiltration scenarios, novel bioweapon uplift, power-seeking generalization. All framed against a "Low risk" conclusion. Classic fear theater.
They do admit real process failures: alignment-faking transcripts contaminating production training data, classifiers disabled for months, unauthorized access gaps, saturated evals that can't track capability growth. Then they still grade themselves as "adequately mitigated."
The pattern is clear: amplify existential-risk framing for regulatory positioning, downplay specific risks when models face real restrictions. Long on speculative threats, short on independent verification, heavy on redactions, entirely self-assessed.
The EA/LessWrong cult talking points are all there, but the room has moved on. This style of AI marketing doesn't land anymore.