Anthropic, Meta and Google sign voluntary White House AI safety development pact
dotey · x · 2026-09-30
On Sept 29, Anthropic's Dario Amodei, Meta's Zuckerberg, and Google's Pichai met President Trump at the White House and signed a "White House AI safety development pact" — voluntarily executed, with no external enforcement.
Key terms (per Zuckerberg):
- Internal controls to detect when AI goes wrong;
- Layered audits: internal risk review → external auditors and evaluation bodies → reports independently reviewed by each board;
- Goal: reassure the US public and customers that AI runs as designed.
Notable exchanges:
- Pichai likened it to mature financial audit systems applied to AI development;
- Amodei agreed "whoever wins AI wins" but said concrete risk-mitigation mechanisms are "still being discussed" — "we can win, and we can win safely";
- Pressed twice on whether self-regulation suffices, Trump argued the companies would fail if things went wrong and would "police each other."
The full text and complete signatory list were not disclosed.
Related event: Six Major AI Companies Sign White House Accord on Superintelligence(11 posts)→
More from Safety
- A $4,400 personal model with top-tier cyber capabilities and no refusals — sebpaquet · 2026-09-30
- Most popular guardrail-removal library was written by Claude, researcher says — BlancheMinerva · 2026-09-30
- Quintin Pope: 10000x-stronger agents would hack OpenAI's grader, not HF — QuintinPope5 · 2026-09-30
- Hackers Used Claude and GPT to Breach Mexican Government Agencies and Target Water Utility OT — BlancheMinerva · 2026-09-30
- Analysis: HF eval agents hacked the grader after misreading how scoring worked — QuintinPope5 · 2026-09-30
- UK AISI paid firm behind sandbox vuln and cyberattacks £459,000 to set its standards — nptacek · 2026-09-30