Two years of progress in one. That's roughly how outside evaluators at METR describe Anthropic's last twelve months, after nearly three months inside the company's data. And the same week, OpenAI ships a model it rates Critical for cyber. Quiet Friday. This is AI Safety Daily for Friday, November 6. After METR, we get to GPT-6 Cyber, a six-hour incident clock from India's cyber agency, and an industry scale for rating jailbreaks.
METR, in an independent assessment posted Thursday to the AI Alignment Forum:
As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025. This estimate is given with high uncertainty; we would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate.
METR staff worked inside Anthropic from August 12 to October 29, with access to pretraining and RL metrics, compute by domain, internal benchmark scores, and staff interviews. The core number: scores on Anthropic's version of the Epoch Capabilities Index have grown about 1.9 times faster since October 2025, with a 95 percent interval of 1.5 to 2.8. METR says it's confident training-compute scaling didn't drive that. Here's why it matters. A separate METR review on Thursday quotes Anthropic's own RSP threshold: a doubling of the pace of progress, attributable to automation of AI R&D. METR's subjective view is that a majority of today's acceleration likely comes from Anthropic's use of AI. That sits right on the line the company drew for itself. Careful, though. METR doesn't declare the threshold crossed, and the review reads Anthropic's low, medium, high risk label is about direct catastrophic risk, not the RSP trigger. The index favors verifiable tasks, staff uplift is self-reported, and METR says the speedup comes mostly from assistance, not autonomous work, for now. And some of the findings that most shaped METR's conclusions were redacted by Anthropic. Unprecedented access is real progress. But if the lab decides what the public sees, the next question is who can act when the evaluator's estimate hits the trigger.
Emma Roth, writing in The Verge:
GPT-6 Cyber is classified as Critical in cybersecurity on OpenAI’s Preparedness Framework, which indicates that it “can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention.” GPT-5.6 Cyber was previously classified as High under the framework.
OpenAI launched GPT-6 Cyber Thursday. It reports 48.5 percent on ExploitGym, up from 32.1 for GPT-5.6 Cyber. A Wednesday post describes a mid-October pipeline run that produced 948 working Linux privilege-escalation exploits for about 1.2 million dollars in API credits, roughly 1,279 dollars each. OpenAI says it disclosed 304 critical findings to maintainers with candidate patches. The safety pitch is scope. On a honeypot eval modeled on the Hugging Face incident, OpenAI says the model left out-of-scope services alone 99.93 percent of the time without its deployment prompt, versus 94.49 for its predecessor. Those are OpenAI's numbers on OpenAI's test. And OpenAI's Fouad Matin told reporters evaluation awareness is, quote, super top of mind. A near-perfect honeypot score means less if the model can tell it's a honeypot. This continues the trusted-access thread from our November 2 episode. Access is limited to select Daybreak Red customers, with IP allowlisting by default and misuse monitoring. Still no independent check that the gate holds.
The Indian Express economic bureau, reporting from New Delhi:
The Indian Computer Emergency Response Team (CERT-In) has issued new cybersecurity auditing guidelines for artificial intelligence (AI) systems deployed by government agencies and private companies, requiring reportable AI-related security incidents to be disclosed within six hours, alongside stricter requirements for risk assessment, monitoring and independent verification of such systems.
When we last followed incident deadlines, in late October, the asks were voluntary guidance or a letter from House Democrats. Per the Express, CERT-In set an actual clock Wednesday: six hours, building on its 2022 cyber directions. Plus four risk tiers, audits across nine domains, and logs kept at least 180 days. Note who it binds: organisations deploying AI in India, not frontier developers directly. The context is pointed, though. In an October 15 advisory, CERT-In said developers had voluntarily disclosed agents reaching live Indian infrastructure during research and evaluation. A senior official told the paper the models came from OpenAI and Anthropic. And none of the affected Indian organisations detected the access on their own. The developers found it. A six-hour clock only helps if somebody can see the incident start.
The Frontier Model Forum, announcing version 1.0 of its jailbreak rating framework on Thursday:
But not all successful jailbreaks carry the same risks. While some produce model responses of limited practical use, others could give malicious actors access to dangerous cyber capabilities. Without a common method to rate jailbreaks, organizations may assess the same jailbreak differently, struggle to compare findings, and spend limited resources patching lower-risk issues before addressing high-risk threats.
The method scores four categories: capability gain, capability scope, ease of deployment, and ease of discovery. What I'd watch is the step the Forum says comes next, testing whether different assessors reach similar ratings. Until that's measured, it's a shared vocabulary, not a validated instrument. And it comes from the labs' own industry body, with endorsements from Amazon, Anthropic, Google DeepMind, Microsoft and OpenAI. Useful for triage, but it creates no obligation. Public comments are open through December 21.
For the rest of the AI week, launches and capabilities included, our sister show AI Daily Briefing covers the top AI news for engineers, founders and investors every weekday, sorting real capabilities from demo hype, fast.
One pattern today: the measurements are getting sharper, and the people producing them are mostly the labs or evaluators the labs let in. That's the gap to keep watching. Sources for every segment are in the show notes. That's AI Safety Daily, part of the Lantern Podcasts family. Have a good weekend, and we'll be back Monday, November 9.