Anthropic Expands AI Agent Security Access and Launches Critical Infrastructure Defense Program

Anthropic published two Security-lane announcements in the same week for the second time in this report’s 41-issue history, the first being 2026-w36. On October 6, Anthropic expanded the Cyber Verification ProgramAnthropic's tiered access system giving vetted security professionals reduced classifier blocks on frontier Claude models for defensive and offensive cyber work. (CVP — its tiered access system giving vetted security professionals reduced classifier blocks on frontier Claude models) from a single tier to three: Defense Access for defensive security teams, Red Team Access for authorized penetration testers, and Specialized Access for safety-critical systems such as power grids and interbank infrastructure ¹. Project GlasswingAnthropic's restricted program that ran from April to October 2026, giving a set of founding partners access to Claude Mythos for security research., the restricted program that found over 129,000 verified vulnerabilities across April to October 2026 with a small cohort of founding partners, was merged into CVP’s Specialized Access tier ¹. CyScenarioBenchAn evaluation that measures whether AI models can plan and execute multi-stage cyber operations under realistic constraints. (an evaluation measuring whether AI models can plan and execute multi-stage cyber operations under realistic constraints) showed Defense Access blocked 46 of 50 offensive tasks while Red Team Access blocked none, providing empirical separation between tiers ¹. Two days later, Anthropic launched the Critical Infrastructure Defense ProgramAnthropic's program embedding Claude models and on-site engineers with security firms protecting power grids, water systems, and transportation networks. (CIDP), embedding Claude models and on-site engineers with 11 founding security partners including CrowdStrike, Dragos, Palo Alto Networks, and Rockwell Automation ².

The same October 8 announcement launched OSS ScannerAnthropic's free opt-in service that delivers automated vulnerability reports from frontier Claude models to open-source project maintainers. (Anthropic’s free opt-in service delivering automated vulnerability reports to open-source project maintainers), funded by the Defender Advantage Fund ². Both announcements extend the security infrastructure pattern Anthropic established in 2026-w36, when it disclosed Claude alignment failures and launched Enterprise Frontier Safeguards, and in 2026-w38, when the Life Sciences Verification Program opened Mythos to credentialed bio-research organizations. The Accenture embedded-evaluation partnership (2026-w38) gave Accenture employee-level access to pre-release Claude models; CIDP gives 11 OT security firms access to Claude plus on-site Anthropic engineers. Where prior programs expanded access to specific high-risk domains — life sciences, enterprise monitoring — the CVP expansion removes the cohort restriction entirely. Any vetted security team can now apply for frontier model access at the tier matching their authorization scope.

Anthropic and Google both moved on the AEO layer this week. Claude Haiku 5.5 launched October 7 at 75% lower cost than Haiku 4.5, with Anthropic simultaneously adding computer use and browser use support to its Python and TypeScript SDKs in beta ³. On OSWorld 2.1, Haiku 5.5 scored 72.4% on the computer-use benchmark, compared to GPT-6 Luna at 48.9% ³. Asana reported 30% latency reduction and up to 2.5x faster inference per agent turn in early testing ³. Anthropic also halved Sonnet 5.5 cache read prices, cutting most agentic task costs by roughly 20% ³. The browser use SDK directly enables AI shopping agents to navigate merchant pages without custom tooling. On the same date, Google Cloud announced the Gemini agent (a universal enterprise agent for work) in private preview at its Gemini at Work 2026 event ⁴. The Gemini agent includes built-in spending limits that pause it when exceeded, plus full audit trails — governance primitives that mirror what agentic checkout systems require before granting purchase authority.

OpenAI expanded ChatGPT Ads’ measurement infrastructure on October 5, naming 20 third-party measurement partners covering attribution, incrementality, and brand safety ⁵. The OAIQ pixelOpenAI's site-side event tracking pixel for measuring conversions and ad performance within ChatGPT Ads campaigns. (OpenAI’s site-side event tracking pixel for ChatGPT Ads) and Conversions API are now live alongside integrations with AppsFlyer, Adjust, Kochava, Haus, DoubleVerify, and Integral Ad Science ⁵. Advanced conversion optimisation now learns from view-through as well as click-through conversions. This is the first AI conversation surface to connect to a full third-party measurement ecosystem at this scale, placing ChatGPT Ads on the same attribution infrastructure advertisers use for Google and Meta. For agentic commerce, that matters at the structural level: the discovery surface that precedes an agentic checkout interaction is now measurable by the same tools that track any other paid channel. Advertising spend toward AI surfaces will follow measurement comparability.