AI safety testing is broken on all three fronts: labs, paid auditors and nonprofits all face warped incentives
joshua_saxe · x · 2026-09-11
Joshua Saxe argues the AI industry's safety testing regime is broken: labs self-test and report non-peer-reviewably, for-profit third parties depend on lab access for their valuations, and labs cherry-pick nonprofits whose reputations hinge on that access. Reacting to ccatalini's point that credible evidence of the agent capability-safety gap sits inside the labs, he calls this a prime place for policymakers to step in.
More from Safety
- Polymarket odds for strict US AI regulation surge to 27% amid extinction warnings — Polymarket · 2026-09-11
- Researcher: Keeping CoT Is Good as Extra Safety Layer, but Debate Lacks Nuance — joshua_saxe · 2026-09-11
- Wiz launches Cyber Arena: 300+ real offensive-security challenges benchmark AI hacking ability — evilsocket · 2026-09-11
- LG Smart TVs reportedly scanned home networks and recorded conversations even offline — Polymarket · 2026-09-11
- DOJ scrutinizes Nvidia's ~$20B Groq licensing deal over merger-review evasion — eyishazyer · 2026-09-11
- IonQ Paper: Breaking ECC 256 Needs 19,397 Qubits and 26 Days of Runtime — jamestagg · 2026-09-11