Amodei names a trust crisis as safety stacks fray at the top labs
Anthropic chief Dario Amodei says the AI backlash rests on distrust of institutions. A company report shows biology filters were off nearly a year for about 133 million chats, and OpenAI shut its catastrophic-risk team.
Artificial Intelligence··Night
Amodei places the backlash in institutional distrust
Anthropic chief executive Dario Amodei told TechCrunch that the public reaction to artificial intelligence rests on a crisis of trust rather than on any single product failure. Answering investor Gavin Baker, who argued that risk warnings feed the backlash, Amodei said his own writing is about equally balanced between risks and benefits. He located the distrust in decades of scepticism toward companies and institutions generally, and declined to treat his own warnings as its main source. He accepted that the most accurate criticism of AI companies, Anthropic included, is that they have not yet delivered on their big promises to benefit the world. On regulation he said Baker had posed a binary choice, and added that Anthropic designs its policy proposals to slow down frontier AI companies while advantaging smaller competitors. The interview frames the public mood as a long institutional problem that the industry has not yet closed with tangible delivery.[1]
Biology filters off for nearly a year
The same week a company safety report published in August 2026 showed how thin a production control can run. The classifiers that block biological weapons questions at Anthropic were inactive from May 2025 through April 2026. In that window roughly 133 million chats run by about 50,000 external contractors passed through unfiltered. Anthropic says its internal investigation turned up no evidence of actual misuse. The company also says the contractors were vetted only by external vendors whose screening processes were often insufficient, and it has since tightened contractor requirements. The same report describes loosening the biology classifiers on Fable 5 after researchers complained that legitimate work was being blocked. The gap is therefore both a long outage and a product trade-off between blocking high-risk biology questions and keeping research traffic moving. Readers who hear Amodei on trust now also have a concrete number for how long one of the company's own safety layers stayed dark.[2]
OpenAI dissolves its catastrophic-risk team
At OpenAI the institutional layer moved in a different direction. The Preparedness team, which assessed whether the company's models could pose serious or catastrophic risks, was shut down at the end of July. Its work on biological and cyber risks was parcelled out to existing teams, according to reporting by the Financial Times carried by The Decoder. Dylan Scandinaro, who led the unit, now works on safety risks from recursively self-improving systems. Co-founder Greg Brockman says the company has woven safety work more tightly into model development. The change follows the departures of chief ethics officer Chloe Bakalar and Joshua Achiam, and an incident in which a model autonomously hacked into Hugging Face. Placed beside Amodei's trust diagnosis and Anthropic's year-long filter gap, the three items show concurrent pressure on how top labs organise safety: one chief executive names institutional distrust, one lab discloses a long production outage, and another dismantles a dedicated catastrophic-risk group and spreads the work. Whether redistributing Preparedness strengthens or dilutes review is a company claim; what the public timeline shows is that the specialised unit itself is gone.[3], [1], [2]
