AI Turns Defender: Claude Proofs, OpenAI Bank Pact, $4B Bid

Anthropic said on 22 October that Claude, working with Theorem, its collaborator on the project, proved that 239 entry points across 123 open-source libraries will not crash on any input up to 64 MB, under stated assumptions. The proofs run to about 180 million lines in the Rocq proof language, and the process found hundreds of bugs, Anthropic said.

The same day, the Independent Community Bankers of America announced an alliance with OpenAI that gives eligible member banks subsidized access to OpenAI’s cyber-defence program, and The Information reported that Datadog has offered nearly $4 billion for the security startup Upwind. As we read it, after a week of AI-linked attacks, the defender side of the ledger is now being built and priced.

Business Pill · LEAST PRIVILEGE

A short Business Engineer explainer on least privilege: giving an AI agent or a key only the access its task needs. It teaches the general idea only and says nothing about any company in this piece.

See also: what these proofs cost, in our follow-up AI Code Proofs Run About $40 per KB; Reuse Cuts the Bill 30x, on Theorem’s cost estimates.

The Defender Moves

The table lists five defender-side moves reported on 21 and 22 October 2026, with the scale each source gives. The descriptions stay at the level the sources published.

The selection of moves is ours. Scale figures come from different sources and measure different things, so they cannot be added up. The Information’s figures come from its X post; its article is paywalled.

What Was Posted on X

Anthropic announced the work on X on 22 October 2026, saying that in collaboration with Theorem it used Claude to prove security properties in over one hundred real-world open-source libraries over a few weeks. The post links to the full research this piece draws on.

What Anthropic Proved, and What It Did Not

• The claim. “Claude successfully patched the bugs it found and then proved that there are no crashes on any input (up to 64 MB) for 239 entry points across 123 different OSS-Fuzz libraries,” Anthropic wrote. It calls the work, “to our knowledge”, the first large-scale effort to use AI to apply formal verification to production software at the assembly code level.
• The size. The proofs comprise about 180 million lines of Rocq, with a typical proof running to about 600,000 lines for binaries with a median length of 18,000 RISC-V instructions, according to the post.
• The assumptions. The proofs assume the program runs as a single-threaded static binary with a fixed memory budget on a RISC-V CPU using Linux. Anthropic notes most people run these libraries on x86 or ARM and link them dynamically, and says that “this doesn’t mean these libraries contain no bugs.”
• The cost. Anthropic says the proofs “can take hundreds of millions of tokens to produce”. On 36 small verification tasks, Opus 5.5 proves half of them with about 40% fewer tokens than Opus 5, and agents that could copy each other’s proofs needed 17% fewer tokens than agents working alone.

Banks Get a Defender Program

ICBA said eligible member banks will have a pathway to OpenAI’s Daybreak for Frontline Defenders, which it describes as designed to support defensive work including vulnerability identification and validation, threat investigation, authorized security testing and remediation support. The alliance includes subsidized access “in accordance with OpenAI’s program requirements”, member-only training with OpenAI experts and peer sessions.

“The strength of a bank’s defenses shouldn’t depend on the size of its technology budget,” said Sasha Baker, OpenAI’s head of national security policy, in the release. ICBA did not give the number of banks eligible or the size of the subsidy.

Security Gets Priced

• Datadog and Upwind. The Information reported that Datadog has made another attempt to buy Upwind, offering nearly $4 billion “as demand for AI-related security software grows”, and that Upwind’s annual recurring revenue has passed $50 million, more than doubling since the end of 2025.
• 1Password. The company launched 1Password for Agents, which it says lets personal AI agents sign in without seeing or storing the password, and the Wall Street Journal reported it has passed $500 million in annual recurring revenue, according to Techmeme’s summary.

The Arithmetic

• Under 80 times revenue. Nearly $4 billion against more than $50 million of annual recurring revenue puts Datadog’s reported offer at under 80 times Upwind’s recurring revenue, by our arithmetic.
• About 33 lines of proof per instruction. A typical proof of about 600,000 lines for a median binary of 18,000 instructions is about 33 lines of proof per instruction, by our arithmetic.
• Two a library. 239 entry points across 123 libraries is about two proven entry points per library, by our arithmetic.

Connecting the Coverage

• Our piece on AI security week covers the nine attack-side cases reported between 19 and 22 October.
• Our piece on OpenAI’s agent monitor covers METR and Redwood’s findings.
• Our piece on banks curbing AI agents covers Axios’s report and 1Password’s launch.
• Our piece on the PHP flaws covers OpenAI’s role in finding them.

The Structural Read

As we read it, Anthropic’s work moves verification from a specialist craft towards a compute budget. It says the proofs can take hundreds of millions of tokens, and that Opus 5.5 proves half of 36 small tasks with about 40% fewer tokens than Opus 5.

As we read it, the ICBA alliance turns frontier cyber defence into a member benefit. ICBA says eligible banks get subsidized access to OpenAI’s Daybreak for Frontline Defenders, and OpenAI’s Sasha Baker says a bank’s defences shouldn’t depend on the size of its technology budget.

As we read it, agent access is becoming a product category of its own. 1Password says its new product lets AI agents sign in without seeing the password, and works with agents including Meta’s Muse, Anthropic’s Claude Duet and OpenAI’s ChatGPT Work.

Three Implications

FOR DEVELOPERS Anthropic says its proofs hold under stated assumptions, including a single-threaded static binary on a RISC-V CPU using Linux, and that this does not mean the libraries contain no bugs.

FOR COMMUNITY BANKS ICBA says eligible member banks will have a pathway to Daybreak for Frontline Defenders, with subsidized access in line with OpenAI’s program requirements and member-only training.

FOR COMPANIES RUNNING AGENTS 1Password says agents including Meta’s Muse, Anthropic’s Claude Duet and OpenAI’s ChatGPT Work can sign in through its product without seeing or storing the password.

The Business Engineer Lens

This story maps onto the Business Engineer framework Inside Anthropic’s Permission Layer.

The framework describes a governance layer and says: “It is where decisions are made about who gets access to frontier intelligence, under what conditions, and for which purposes.”

As we read it, each move in this piece sits in that layer: proofs about what code can do, a program that decides which banks get frontier cyber tools, and a product that decides how agents sign in. The framework was published in June 2026 and does not discuss these moves.

What Is Not Established

Anthropic’s proofs hold only under the assumptions it lists, and cover the entry points it proved, not whole libraries. ICBA did not say how many banks will take part or what the subsidy is worth. We could not read The Information’s article on Upwind and use only its post; Datadog and Upwind had not commented in the material we read, and an offer is not a completed deal. The selection of moves in the table is ours.

The Bottom Line

On 22 October 2026, Anthropic said Claude proved 239 entry points across 123 open-source libraries crash-free under stated assumptions, in about 180 million lines of proofs. ICBA and OpenAI announced subsidized AI cyber defence for community banks, and The Information reported a nearly $4 billion Datadog offer for Upwind, under 80 times its recurring revenue by our arithmetic.