Friday, November 6, 2026
TL;DR
METR's independent review says Anthropic's own safety report understates how close its models already are to AI-accelerated self-improvement - plausibly already past the threshold Anthropic's IPO filing called merely imminent. The same week, a historic tech selloff deepened investor doubt that Anthropic's IPO can even clear its already-reduced valuation. Separately, OpenAI's new cyber model and the open-weight GLM-5.3 each independently built working Firefox exploits, days apart.
Act on this
- Update Firefox now: two unrelated AI systems, one closed and one open-weight, each built working exploit chains against it this week. CyberScoop, OpenAI launches GPT-6 Cyber #
Signals
Finding and chaining a working exploit from a vendor's own public patch notes is now routine on both sides of the open-weight/closed divide, not one lab's edge. #
OpenAI's new GPT-6 Cyber, launched Nov 5 to restricted customers, found 81 novel Firefox bugs and a critical Brotli RCE. Days earlier, Epoch AI showed the open-weight GLM-5.3 exploiting six straight Firefox releases in as little as 12 hours each, per Zvi Mowshowitz's Nov 5 roundup. CyberScoop, OpenAI launches GPT-6 Cyber · Zvi Mowshowitz, AI #193: Progress May Dramatically Accelerate
Anthropic's own account of how fast AI is already accelerating its research undersells it, according to the outside evaluator with direct access to the data. #
METR's Nov 5 review of Anthropic's own August Risk Report says the public version understates AI-driven R&D acceleration, calling it plausible models already crossed the Responsible Scaling Policy's 2x threshold Anthropic's own IPO filing called merely 'imminent.' METR, R&D section review of Anthropic's Risk Report
News
WATCHNasdaq posts its steepest drop since 2020 as AI-spending and election fears collide #
Tech stocks lost well over $2 trillion in value as investors priced in both AI-capex doubts and the prospect of tougher Democratic oversight after the party's midterm wins; chip stocks led the decline. WSJ, Wall Street Gets $2.5 Trillion AI Reality Check
WATCHGoogle commits $17 billion to a new Wyoming data center, pledges to cover its power costs #
Google says it will pay for all electricity the site uses and is forming a community council for local input - a pitch for buy-in after voters rejected data centers in 40 of 41 ballot measures nationwide last week. Bloomberg, Google plans to spend $17 billion on Wyoming data center
WATCHMicrosoft closes its first US bond sale since 2017 as Meta seeks up to $20 billion more #
Microsoft's $25 billion offering drew about $115 billion in orders and closed Nov 4; Meta is separately marketing at least $15 billion in new bonds after S&P cut its outlook to negative over AI capex and cash burn. TheNextWeb, Microsoft sells at least $25bn of bonds · Dealroom, Meta borrows $15B as investors demand more
SHIPThinking Machines Lab previews a 594-billion-parameter open-weight model #
Inkling 2 Small, a mixture-of-experts model with 17B active parameters, ties or beats Gemini 3.8 Flash on several benchmarks. Full Apache 2.0 weights are promised next week after external safety testing; only API access is live now. Thinking Machines Lab, Inkling 2 Small
The long view
A year ago, AI speeding up its own research was a line in Anthropic's safety policy naming a future threshold - no company claimed proximity, and no outside group checked a lab's self-assessment against the data behind it. On Nov 5, METR, given exactly that access under an October program, said Anthropic's own public account understates the acceleration, calling it plausible the threshold was already crossed by July. If this holds, outside evaluators start checking labs' safety disclosures against internal data in near-real time, not years later.
Also noted
- TeraWulf's Nov 4 SEC filing shows Google will backstop Anthropic's $19 billion Kentucky lease obligations if Anthropic defaults or goes insolvent. TeraWulf, Form 8-K #