AI research wiki
responsible-ai · 2026-11-06 · Tier 2

GPT-6 Cyber ships, and METR measures AI R&D acceleration at Anthropic

GPT-6 Cyber ships, and METR measures AI R&D acceleration at Anthropic

Sources: Introducing GPT-6 Cyber, OpenAI (page blocked to automated fetch, figures via The Decoder) · The Decoder · @OpenAI · METR assessment · The Information · raw: Decoder, lab watcher

TL;DR. Two disclosures on how fast capability is moving. GPT-6 Cyber is OpenAI's specialized security model, the one behind the 948 Linux exploits reported on 11-05. It is rated Critical under OpenAI's Preparedness Framework and limited to select Daybreak Red customers. METR, after an 11-week review of Anthropic's internal data, estimates model progress has accelerated about 1.9x, and guesses roughly 80% of that comes from Anthropic's own use of AI. Anthropic redacted most of the evidence METR found most informative.

GPT-6 Cyber

  • Benchmarks (all from OpenAI, no independent tests yet): ExploitGym 48.5% vs 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra on Daybreak Blue (0.0% with safeguards). A new web-app test without source access: 40.3% vs 32.0% for Astra.
  • Real software: 1,724 Linux kernel vulnerabilities reproduced, 948 working exploits, against 166 for GPT-5.6 Sol under the same conditions. 81 previously unknown critical Firefox vulnerabilities, and 30 known ones found more severe than thought. Mozilla shipped fixes.
  • Campaign cost: $1,213,430 in compute, about $1,279 per exploit. 304 critical findings reported to Linux maintainers with fixes.
  • Staying in scope: trained with a honeypot (ExploitGym Honeypot plants easy-to-hack side services that also reveal the task flag). It ignored the bait in 99.93% of tests without a deployment prompt, against 94.49% for its predecessor, and in 100% with it.
  • Access: select Daybreak Red customers via API and Codex, expanding over time. Partners named: Mozilla, Palo Alto Networks, Calif, Cisco. A Cyber at the Frontier livestream is set for 11-12. OpenAI also announced a JPMorganChase partnership with a security focus.

METR on Anthropic

  • AECI benchmark scores accelerated 1.9x (95% CI 1.5-2.8x) since October 2025, with robustness analyses from 1.1x to 2.9x. METR's reading is "likely around twice the rate" of March 2024 to October 2025. Non-Anthropic frontier models show 1.4x.
  • METR guesses "perhaps ≈80%" of current acceleration is due to AI, with substantial uncertainty, and says it was not from faster growth in training compute.
  • No convincing bottleneck in algorithms, training compute, experiment compute or data in the next 12 months, stated with low confidence.
  • Self-reported staff uplift is about 12x, which METR calls very loose evidence.
  • The redactions matter. Anthropic held unilateral redaction rights and "redacted a majority of the evidence we found most informative," saying the aim was removing information that could help rivals. It canceled a compute-allocation check (capabilities vs safety) midway. METR used free tokens and Anthropic devices and took no payment.
  • Reaction: Casey Newton and others highlighted the redactions, and an Anthropic cofounder's defense of them drew criticism in the feed.

How this relates to prior wiki pages

  • Resolves a same-week report: The Information via the 11-05 digest said a cyber model could land "as soon as Friday." It shipped Thursday.
  • Confirms the same page's report that Anthropic's S-1 says it "may imminently meet" its RSP threshold for dramatically accelerated AI-driven progress. METR now puts a third-party number under that language.
  • Contrast: the 11-05 finding that GPT-6 Astra quietly did about 40% worse on safety research shows how little a capability number says about effort on safety work. Cyber's 99.93% scope score is the opposite kind of claim, a measured behavior with a clear oracle.
  • Touches agent harness engineering: Daybreak's tiering is routing by verified identity, not by task.

Gaps

OpenAI's numbers lack independent replication, and "novel technique" claims are disputed by researchers. Honeypot tests measure a specific bait, not general scope discipline. METR's central estimates rest on evidence the public cannot inspect, and its report says transparency and redaction-authority items on its own checklist were not met.