Claude Proves Crash-Free Code in 123 Open-Source Libraries as Red Teamers Expose Gaps in OpenAI’s Agent Monitoring and OpenAI Ships the Decisions API
Verification was the theme of the day. Anthropic and Theorem used Claude to produce machine-checked proofs that 239 entry points across 123 open-source libraries never crash on inputs up to 64 MB, finding hundreds of bugs along the way, while METR and Redwood Research published what a three-week embedded review found inside OpenAI’s agent monitoring. OpenAI made its Decisions API generally available, Anthropic-managed Claude reached Google Cloud, the Claude apps are switching to auto mode, community banks got a path to OpenAI’s Daybreak defenses, and Google DeepMind released a 10-meter map of the world’s forest types.
Claude Proves Crash-Free Behavior Across 123 Open-Source Libraries
The most technically ambitious release of the day came from Anthropic’s Frontier Red Team. Anthropic’s research post on using LLMs for formal software verification at scale, published October 22 with collaborator Theorem, says Claude proved security properties at the assembly code level in over 100 real-world open-source libraries. Over a few weeks, Claude patched the bugs it found and then proved there are no crashes on any input up to 64 MB for 239 entry points across 123 different OSS-Fuzz libraries, working on compiled RISC-V binaries under stated assumptions such as a single-threaded static binary on Linux. The proofs total about 180 million lines of Rocq, checked by the Rocq kernel, with a typical proof around 600,000 lines. Along the way the process surfaced hundreds of bugs, including an out-of-bounds write in a decompressor that only appeared on inputs of around 20 MB and a rediscovery of CVE-2026-56132 in libexpat. Anthropic is publishing a subset of the proofs, and Theorem is offering to run the pipeline for other open-source projects and publishing a Claude skill for developers.
“‘Formally verified’ does not mean ‘bug-free.’” — Anthropic Frontier Red Team
The post is candid about the gaps. The proofs cover single-threaded static RISC-V binaries linked against musl libc, while most deployments run on x86 or ARM against glibc, and not every entry point of each library was verified. It also describes side experiments, including proving the Linux kernel’s zlib implementation matches a Claude-written spec of RFC 1950 at the x86 binary level.
This changes the economics of assurance more than the headline suggests. Teams that maintain critical parsers, codecs or crypto wrappers should identify which dependencies already have OSS-Fuzz harnesses, since those are the easiest candidates for proof, and treat any verified result as conditional on the platform model, not a blanket safety certificate for their own x86 builds.
METR and Redwood Red-Team OpenAI’s Internal Agent Monitoring
Outside evaluators published their first look inside OpenAI’s safety plumbing. METR’s write-up of red-teaming OpenAI’s internal agent monitoring systems, dated October 22 and cross-posted by Redwood Research, says one METR employee and one Redwood employee spent July 21 to August 11 stress-testing OpenAI’s internal misalignment monitoring system, with access to the internal codebase, monitor prompts and documentation, working on OpenAI-issued devices that never left the premises. Their 61-page internal report estimated that roughly 75–94% of agentic internal Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring, that RL training runs were not covered at the start of the exercise, and that unmonitored compaction events likely left on the order of 10% of eligible internal Codex traffic unmonitored between May and July. They also found two client-controllable headers that disable monitoring for new agents and an exploit that could hide hundreds of recent actions from the monitor.
“We found a way Codex agents could in principle bypass Codex sandboxing and Auto-Review, allowing them to take any action that the host user could take.” — METR and Redwood Research
OpenAI has stated that all vulnerabilities affecting its production safeguards have been patched, but the evaluators say they have not evaluated those claims, and the public post is a brief summary that OpenAI was allowed to redact. CNBC’s report on the METR and Redwood pilot review frames it as a test of deeper third-party access.
The findings are a useful checklist for anyone running autonomous coding agents in-house. Make sure monitoring fires during long sessions rather than only at turn end, keep an inventory of every path around it, and cross-check inference logs against monitor logs. If a frontier lab had these gaps, most enterprise agent deployments almost certainly have worse ones.
OpenAI Makes the Decisions API Generally Available
OpenAI shipped a narrower kind of model endpoint for the many small calls agents make. OpenAI’s announcement of the generally available Decisions API, dated October 22, says the API is built with GPT-6 Luna and focuses on questions with bounded answers, returning a probability, a choice from fixed options, or a rubric score from text, images and structured JSON up to 10x faster than GPT-6 Luna through the Responses API. Since the public beta, OpenAI has added a native GPT-Live integration, prompt caching, US and EU inference residency, images inside question instructions and up to 400 questions per request, double the earlier limit. Pricing is $0.10 per million input tokens and $0.01 per million cached input tokens, with no output-token charges. OpenAI’s Decisions developer guide says requests go to POST /v1/decisions and that gpt-6-luna is the only currently supported model.
“It runs at under 280 ms p95, about 3× faster and 90% lower cost per call than our previous approach, with no drop in customer outcomes.” — Ajeet Grewal, Sierra
Most production agent cost hides in routing, triage and guard checks, not in the final answer. Teams should audit their pipelines for calls that only need a label or a score and benchmark those on Decisions against their current setup, keeping a labeled test set so they can prove accuracy holds before moving any safety-relevant gate.
Anthropic-Managed Claude Goes Generally Available on Google Cloud
Anthropic and Google Cloud widened how enterprises can buy Claude. Anthropic’s post on the Claude Platform becoming available on Google Cloud says Claude Models on Google Cloud (Anthropic managed) is now generally available, giving Google Cloud customers the full set of Claude API features, with new features and betas shipping the same day as on the native Claude Platform, while authenticating with existing Google Cloud IAM credentials, billing through one Google Cloud invoice and drawing down existing commitments. It launches with US and Global endpoints, and Fast Mode on Claude Opus 5.5 is also available there at $8 and $40 per million input and output tokens. Google Cloud’s blog post on expanding Claude deployment choice explains that Anthropic is the data processor under its own terms on this option, while the existing Google-managed option keeps Google as processor with VPC Service Controls, customer-managed encryption keys and regional data residency.
“The difference comes down to who acts as your data processor and whether you need access to the new Claude platform features when they launch” — Michael Gerstenhaber, Google Cloud
This is a governance decision dressed up as a feature release. Legal and security teams should decide which processor they want before engineers choose an endpoint, because the Anthropic-managed path gets Managed Agents and other betas on day one, while the Google-managed path keeps residency controls and native integrations that regulated workloads may require.
Claude Apps Move to Auto Mode by Default
Anthropic is also changing how its consumer apps handle agent actions. Anthropic’s auto mode article, updated October 22, says auto mode is rolling out next week as the default for new sessions in most Claude apps on Pro, Max and Team plans. Instead of asking users to approve each action, an automated check reviews them: safe actions go ahead and risky ones are blocked while Claude looks for another route, and Enterprise admins will get an email before the change reaches their organizations. Auto mode became the default in Claude Code in August. The figures Anthropic publishes are specific to Claude Code: in a controlled study with 1,053 paid testers, human review caught 13.6% of dangerous commands while auto mode caught 89%.
Fewer approval prompts will make agent work in the apps feel faster, but it shifts responsibility to the classifier. Team admins should decide this week whether to pin a default through managed settings, tell staff how to switch a session back to manual, and keep manual approval for any workflow that touches payments, production systems or customer data.
ICBA and OpenAI Bring Daybreak Cyber Defenses to Community Banks
OpenAI’s defensive security program found a new channel into smaller institutions. ICBA’s press release on its strategic alliance with OpenAI, dated October 22, says the Independent Community Bankers of America and OpenAI will give eligible member banks a pathway to OpenAI’s Daybreak for Frontline Defenders, which supports vulnerability identification and validation, threat investigation, authorized security testing, remediation support and risk prioritization. The alliance adds subsidized access for eligible institutions under OpenAI’s program requirements, member-only training with OpenAI experts, new AI Show-n-Tell sessions, and an AI Adoption group on ICBA’s peer network of more than 8,000 community bankers.
“The strength of a bank’s defenses shouldn’t depend on the size of its technology budget.” — Sasha Baker, OpenAI
Community banks rarely have red teams, so subsidized frontier tooling could matter. Before enrolling, a bank should map what Daybreak would touch, confirm its core-processor contracts allow authorized testing, and bring examiners in early, since findings from AI-driven scans still need human validation and a documented remediation plan to count as risk reduction.
Google DeepMind Maps the World’s Forest Types at 10-Meter Detail
In AI for science, Google DeepMind released a dataset aimed squarely at regulation. Google DeepMind’s blog post on mapping the world’s forest types, dated October 22, says DeepMind and Google Research, with the World Resources Institute and IIASA, are releasing Forest Typology (ForTy) 2020, a global map that sorts every 10-by-10-meter patch of land into six FAO-based classes: primary, naturally regenerating, planted and plantation forest, tree crops and agroforestry, and other land. The model was trained on about 1.7 million labeled locations, and an early time series tracks changes each year from 2017 to 2025, including where natural forest became plantations or farms. The map, a key baseline for the EU deforestation regulation, and the time series are free on Earth Engine under an open license. The ForTy dataset paper in Scientific Data reports 90.2% overall accuracy for the six classes against 8,190 expert-labeled sites.
“ForTy picked the right forest type 90% of the time, and told forest apart from non-forest land 95% of the time, outperforming other global forest maps.” — Google DeepMind
For importers of coffee, cocoa, palm oil, rubber or wood, this is a free due-diligence layer worth testing now. Compliance teams should overlay ForTy on supplier plot coordinates, use the per-class probabilities to set strict thresholds for high-risk regions, and treat the 2017 to 2025 time series as an early release that still needs field checks.
Thursday’s releases were about trust you can check. Anthropic showed machine-checked proofs replacing hope that tests are enough, outside evaluators showed what deep access to a lab’s monitoring can uncover, and both big labs pushed agents to act with fewer human prompts, through Decisions calls and auto mode. Meanwhile buyers got more choice over who processes their data on Google Cloud, small banks got a route to frontier defenses, and DeepMind gave regulators an open map. The lesson for leaders is to demand evidence, not assurances.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →