Anthropic goes public as GAO tests show AI safeguards bypassed
Tracking: Frontier AI model safeguards under lab safety frameworksPolicy matter covering the safety measures frontier AI developers adopt under their own responsible scaling frameworks, including deployment safeguards for CBRN and cyber risks, model-weight security controls, and whether such voluntary standards should be mandated or independently verified.
- Earliest seen
- Nov 5
- METR · Thursday, Nov 5
- Articles · 30d
- 23
- 5.8 per story
- Outlets
- 9
- Most: Anthropic (13)
- Hill · 30d
- 2
- Last: Friday, Nov 20 · none ahead
- Contracts · 30d
- 0
- No awards or solicitations
- Momentum
- 1.5
- Articles ▼1 vs prior week
Activity
30 days back, 14 ahead · hollow = scheduledWhat's happening
Updated yesterdayGAO covert tests found ChatGPT, Claude, Gemini and Grok produced feasible attack plans, often after safeguards were bypassed. Anthropic marked its IPO while promoting third-party evaluation, METR access and SB 53, and METR's reviews say Anthropic's risk report understated acceleration evidence and that key evidence was redacted. No hearing or vote is scheduled in the items.
Latest
YesterdayOur first signal
Hill
2 Hill itemsGAO issued two reports: covert tests showing four major AI platforms gave attack plans, and an AI-cybersecurity report with 16 policy options. Anthropic endorsed California's SB 53 and federal frontier AI safety standards.
Contracts
From news coverageAnthropic and Accenture plan at least $1 billion each over five years for embedded model evaluation, and Anthropic launched a fund for third-party evaluations. It also opened an expanded Cyber Verification Program with reduced-safeguard access for vetted security teams.
Pressure
13 articlesAnthropic's CEO Dario Amodei reaffirmed a pacing pledge at the IPO bell, with Altman, Hassabis and Musk cited as backing his position. METR, forecasters and Axios coverage stress accelerating capabilities and transparency gaps.
Who covered it, when
A dot per article · ringed = earliest in our sourcesForces at work
share of articles- Threats & findings6 · 26%GAO found safeguards bypassed on four major platforms. METR finds AI R&D acceleration of roughly 2x at Anthropic, with redactions, and an exponential capability trend.GAO · METR · Anthropic
- Industry advocacy6 · 26%Anthropic pushes third-party evaluation, standardized red teaming, and voluntary frameworks like its Responsible Scaling Policy, with $1B+ pledges.Anthropic · Accenture
- Regulation4 · 17%Both are shortening coordinated vulnerability disclosure windows toward 30 days as AI speeds exploitation; Anthropic revised its Responsible Scaling Policy to version 3.4 in July.Google Project Zero · Anthropic Frontier Red Team
- Technical milestones2 · 9%Anthropic applied ASL-3 protections to Claude Opus 4 and reports mixed results in cyber competitions.Anthropic
- Oversight2 · 9%METR reviewed Anthropic's Risk Report, saying the 2x threshold was plausibly crossed and acceleration evidence understated; Anthropic gave METR staff employee-like access.METR · Anthropic
- Public opinion1 · 4%Amodei used the IPO day speech to reaffirm pacing AI development until safety is assured.Dario Amodei · Daniela Amodei · Sam Altman · Demis Hassabis · Elon Musk
- Legislation1 · 4%Anthropic endorses SB 53 and federal standards, with reservations on the compute-based threshold.Anthropic
- Procurement & funding1 · 4%Cyber Verification Program offers tiered reduced-safeguard access to vetted defenders, with data retention for misuse monitoring.Anthropic
In the news
23 articlesOn the Hill
22d agoCongressGAO's report on AI and cybersecurity describes how generative AI has boosted both cyberattacks and defensive tools, and lays out 16 policy options to close governance and testing gaps.CyberAIAnthropic goes public as GAO tests show AI safeguards bypassedDetails13 linked 4d agoCongressGAO's covert tests found that four widely used generative AI platforms, including ChatGPT, Claude, Gemini and Grok, produced verified, feasible attack plans for an IED mass-casualty attack and for grid attacks using drones or firearms, often after safeguards were bypassed.CyberAIAnthropic goes public as GAO tests show AI safeguards bypassedDetails35 linked