A thick steel vault door with heavy hinges and locking bolts stands partly open in a bare concrete wall, a dim room of black server racks with small blue lights visible through the opening, a keypad on the wall beside it and a plain concrete floor in front.

The Safety Case

Anthropic Says It May Soon Cross Its Own Line

Anthropic’s amended prospectus says it “may imminently meet” its safety policy’s threshold for AI-accelerated progress; in August it said it might cross it “in the coming year.”

SAN FRANCISCO, Nov. 2 — Anthropic told prospective shareholders on Monday that it may soon cross a line drawn by its own safety policy: the point at which AI, put to work on AI research, dramatically speeds up the progress of AI itself. Its internal estimates and its expectations for future models, “rather than our current models,” the company wrote in an amended prospectus, “indicate that we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI.’”

“We have not designated any of our current models as exceeding this threshold,” the filing says. Neither sentence was in the version Anthropic made public on Oct. 26.

In August the company sounded further off: its risk report published Aug. 14, covering developments through July 15, concluded that “the overall pace of progress remains below the RSP threshold,” called that conclusion “uncertain and subject to some lag,” and said “we may cross this threshold in the coming year.” The prospectus also says what caution costs: precautions such as delayed releases “could slow our capability gains relative to competitors that do not adopt similar constraints.”

The warning comes with an admission. “We have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred,” the amendment says, where the October text said only that it “may not identify” one. It does not say which acceleration was missed.

The policy, in the version in force since July 8, treats the threshold as met if Anthropic’s models could “fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs,” or if there is “‘dramatic acceleration’ of the pace of AI progress for reasons that likely relate to the automation of AI R&D.”The TestThe second test is met where Anthropic observes or expects progress in aggregate AI capabilities at double both the rate it would expect and the fastest extended rate seen without significant AI contributions to AI research, with the doubling plausibly owed in substantial part to automated research and likely to lead to even greater acceleration. Doubling means, the policy says, “as much progress in one year as one would see in two years at baseline.” The threshold, its changelog says, “is intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize.”

At that threshold the policy’s planned mitigations include “moonshot R&D for security” projects and risk reports reviewed by at least one outside expert. One of its commitments uses the filing’s word: where Anthropic has developed “or will imminently develop” such a model, with clear evidence that no competitor will soon develop one, it “will require a strong argument that catastrophic risk is contained” and will delay development and deployment “as needed to achieve this,” until it no longer believes it has a significant lead. The filing does not mention that commitment.

The amendment adds that Anthropic does “not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding,” and that it is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic.” A new risk-factor sentence says it is considering measures, including policy updates, “to require human review and approval of certain decisions in our research and development processes.”

Its figures on Anthropic’s own use of Claude are unchanged from October. The value of internal Claude usage at list prices “has grown by over 250 times from January 2026 to September 2026,” it says, and “over 90% of the contributions to our codebase in September 2026 were written by Claude.”Counting CodeThe filing does not define a contribution. In a June 4 essay, the company’s Anthropic Institute put Claude’s share of code merged into Anthropic’s codebase at “more than 80%” as of May 2026, a count it called more conservative than leadership’s estimate of 90 percent or more. It called lines of code “an imperfect measure.”

A risk factor lists what could go wrong; it is not a finding, and the filing claims no crossing. TechCrunch, which reported the passage on Monday, offered the doubter’s reading: “Skeptics might point out that an S-1 functions as both disclosure and marketing, and that claims of being close to a model so powerful that it can build itself could fall into the latter category.”

What to watch: the policy update and the “further reports” the filing promises, and a further amendment with the price range, still blank.

Daily Cortex is written with AI models, including Anthropic’s Claude, and fact-checked with OpenAI’s GPT models.

Sources

  1. U.S. Securities and Exchange Commission (Anthropic, PBC), “Form S-1/A, Amendment No. 1 to the registration statement, and preliminary prospectus,” Nov. 2, 2026.
  2. U.S. Securities and Exchange Commission (Anthropic, PBC), “Form S-1 registration statement and preliminary prospectus,” Oct. 26, 2026.
  3. Anthropic, “Responsible Scaling Policy, Version 3.4,” July 8, 2026.
  4. Anthropic, “Risk Report: August 2026,” Aug. 14, 2026.
  5. Anthropic, “Anthropic’s Responsible Scaling Policy (versions and updates),” Aug. 14, 2026.
  6. The Anthropic Institute, “When AI builds itself,” June 4, 2026.
  7. TechCrunch, “Anthropic warns that it’s very close to RSI capability, aka self-building AI,” Nov. 2, 2026.