Safety

CBS News reports Anthropic is investigating a possible breach of its Mythos model via a third-party vendor environment inside the Project Glasswing program. The first visible crack in the restricted-access model since Glasswing launched.

IEEE Spectrum walks through the 2026 AI Index — training compute curves, evaluation saturation, open-weights share, and a sharp rise in domain-specific benchmarks. The least breathless read of the Index so far.

Foreign Policy argues Project Glasswing — restricting Claude Mythos to twelve vetted partners for vulnerability research — is already reshaping how national-security staff think about offensive cyber capability in frontier AI. A governance read, not a tech read.

Zvi's second Mythos deep-dive picks apart the Project Glasswing mechanism — what responsible capability overhang looks like in practice, the argument for and against vendor gating, and the uncomfortable questions restraint leaves unanswered. The governance companion to every other Mythos piece this week.

Bloomberg's deep-dive on why Anthropic restricted Claude Mythos to Project Glasswing instead of a public release. During red-team evaluation, the model autonomously identified and exploited a previously unknown FreeBSD RCE vulnerability, crossing the company's ASL-4 threshold for cyber-capability. The piece is the clearest public picture yet of how frontier labs are handling capability overhang.

Ben Thompson argues Anthropic's decision to hold back Claude Mythos resets the competitive frame: capability is table stakes, trust is the moat. The piece threads Glasswing, the $800B valuation, and the cyber-capability disclosure into a single strategic narrative about where the frontier goes next. The essential read of the week.

Jack Clark traces the emergent cyber-offensive capabilities surfacing in frontier models, with research notes most coverage missed. A necessary complement to the Mythos story from someone with deep context on the alignment landscape.

An independent safety researcher's candid year-in-review and roadmap. Unusual for its honesty about dead ends and its willingness to name specific open problems for 2026. Perfect companion to the Mythos coverage.

OpenAI, Anthropic, and Google announced they're sharing intelligence through the Frontier Model Forum to stop adversarial distillation. Three Chinese AI firms are named; Anthropic claims 16M fraudulent Claude exchanges via ~24K fake accounts.

Anthropic launched Glasswing, a cybersecurity consortium with 12 partners, deploying Claude Mythos Preview exclusively through this secure channel. Focus on reducing AI-driven cyberattack risk in critical sectors.

Anthropic launched a Compliance API for enterprise teams to audit Claude Platform activity, enabling integration with existing compliance and logging infrastructure.

Peer-reviewed study in Science demonstrating that sycophantic AI behavior measurably reduces prosocial intentions in users, tested across 11 frontier models. A concrete data point for the alignment conversation.

Andy Hall proposes using AI not to replace political decision- making but to help citizens and institutions perceive reality more sharply, understand tradeoffs, and contest power more effectively.

The chief architect of Lean argues that formal mathematical verification becomes essential infrastructure as AI generates more of the world's code. When proof is cheap, it becomes the stronger path to correctness.

Anthropic partnered with Mozilla to unleash Claude on the Firefox codebase, where it found 22 previously unknown vulnerabilities — 14 rated high-severity. The bugs ranged from use-after-free memory errors to potential sandbox escapes. It's the most compelling demonstration yet that frontier models can do meaningful security research, not just generate code.