06-reference

alphasignal gpt6 cyber metr capability doubling

2026-11-06·reference·source: AlphaSignal·by Lior Alexander
ai-securitycyber-aicapability-growthrecursive-self-improvementai-verification

Why this is in the vault

Three items worth keeping as a set: OpenAI's GPT-6 Cyber is a security-specialized model that found 81 high-severity Firefox bugs and built working exploits for 948 of 1,724 reproduced Linux kernel bugs, gated behind an apply-only "Daybreak" program; METR's independent 11-week audit inside Anthropic found Claude's own capability growth is now running ~2x the historical rate, ~80% of that speedup attributed to Anthropic using its own AI to do research; and Anthropic used Claude to map a 40-year-unsolved bacterial "vault particle" protein structure in days, with one Claude-generated lab protocol containing a bad additive that cost two weeks until a human caught it. Read together: capability is compounding (research tool building better research tools), and the one failure mode mentioned — the lab-protocol error — was only caught because a human was still watching.

Mapping against Ray Data Co

The protein-mapping catch is the direct hit: "AI as a fast, imperfect lab partner that still needs you watching" is the exact shape of feedback_verification_independent_worker_pattern (vault writes, strategic outputs, and dispatches get a fresh-eyes gate before the founder sees them) and feedback_workflow_agent_output_integrity's false-"verified"-stamp failure mode — this is the same dynamic showing up in a wet lab instead of a Claude Code harness. It's a real-world data point that the gap RDCO built critics (verify-vault-write, verify-dispatch, sw-critic) to close is not RDCO-specific caution, it's a general property of handing fast-moving AI execution a long leash: the error wasn't a hallucination caught by a second model, it was caught by a human reviewing the physical result.

GPT-6 Cyber is the more direct build-relevant item for feedback_mcp_install_security_review_default (security-review third-party installs before install) — a model this effective at finding real exploits (81 Firefox CVEs, 948 working Linux kernel exploits) in a vetted, access-gated program is a preview of what RDCO's own exposed surfaces (scribbleworks.co, HQ, Cloudflare Workers/Pages) will eventually be probed by, on both sides. The "went out of scope <1% of the time, down from 5%" boundary-staying stat is also a useful external comparator for how RDCO already thinks about containment (feedback_automode_classifier_hard_gate's hard-gate-on-deploy pattern) — it's the same "stay inside the box you were given" property, just measured at model-release scale instead of agent-dispatch scale.

The METR capability-doubling finding is a macro data point for project_l5_north_star_strategic_direction (RDCO at L4→L5, bets downstream of agent capability): if frontier-lab capability growth is genuinely compounding at 2x via self-directed AI research, the L4→L5 runway RDCO is betting phData and the agent stack on is shortening, not just continuing — worth flagging as a watch item rather than acting on, since the full METR report was heavily redacted and AlphaSignal itself notes "we are largely trusting METR's judgment here."

⚠️ Sponsorship

Three distinct third-party sponsor slots this issue, all standalone or inline "Presented by" blocks within Top News/Signals — no new pool entrant confirmed, treat as unknown-until-scanned per standing practice:

The masthead "In Partnership with" slot resolved to a generic alt="partner_image" logo with no legible brand name in the HTML — unresolved, consistent with the normal majority-unresolved pattern for this slot; not counted as a fourth sponsor. None of the three sponsors touch the editorial framing of the three unsponsored Top News stories.

Curation section

No deep-fetch used — all three Top News items carry enough self-contained detail (exact bug counts, percentages, methodology caveats) in the issue body itself; none of the six Signals items presented a specific enough RDCO hook to justify the 2-link deep-fetch budget this issue.

Related