🧭 Anthropic Launches Cyber Mission — Free Vulnerability Scanner for Open-Source Projects Finds 29,000 Candidate Issues at 88% Precision
Anthropic unveiled its Cyber Mission on October 8 — a long-term defensive security programme anchored by two immediate products. The Critical Infrastructure Defense Programme (CIDP) deploys frontier Claude models and Anthropic on-site engineers to power grids, water utilities, and government networks through founding partners including CrowdStrike, Palo Alto Networks, Dragos, and Deloitte. Alongside it, OSS Scanner launched as a free, opt-in service that runs Claude Mythos over open-source codebases to find security vulnerabilities — an early run across several major projects identified 29,000 candidate vulnerabilities with an 88% verified true-positive rate.
How OSS Scanner works
Enrolment: Open a pull request to red.anthropic.com/oss-scanner listing your GitHub repository. Eligibility mirrors Google OSS-Fuzz criteria — actively maintained, wide-deployment open-source projects.
Scan cycle: Claude Mythos analyses the repository on a rolling basis; findings are reported as GitHub issues with reproduction steps and suggested patches, not just line numbers.
Free remediation credits: Enrolled maintainers receive a complimentary Claude Max subscription to accelerate patch development — the intent is to close the find-and-fix loop rather than just surface a list of alerts.
Zero cost: Both enrolment and ongoing scans are free for qualifying open-source projects. Anthropic's stated goal is increasing the security baseline of the broader software ecosystem, not monetising the scanner directly.
Should you enrol your project?
If your project is public, actively maintained, and widely depended upon — and you'd welcome automated triage with model-generated fix suggestions — enrolment is low risk and high upside. The 88% precision figure means roughly 1 in 8 findings will be false positives; treat the output as a high-quality triage pass rather than a definitive audit. You still need human review before acting on a suggested patch.
🧭 Anthropic Rewrites Usage Policy for 2026 — Expanded Civic Use, New Hardware Safeguards, and Protections for Model Wellbeing
Anthropic published its 2026 Usage Policy overhaul on October 8, effective November 12, 2026. The revision is the most significant since the policy's inception and touches four key areas: election-content rules, autonomous-hardware constraints, weapons language, and — notably — a new clause addressing model treatment. Developers have until November 12 to review their use cases for compliance.
Key changes
Election rules narrowed and clarified: The previous broad ban on "political content" has been replaced with a precise prohibition on deceiving voters or disrupting elections. Legitimate civic uses — multilingual voter information, candidate Q&A tools, get-out-the-vote applications — are now explicitly permitted. This meaningfully expands the design space for civic-tech and journalism products.
Autonomous-hardware safeguards: A new section requires that any physical system operating Claude must maintain an operator override mechanism and implement a fail-safe if Claude disconnects. This is directly actionable for teams building robotics, industrial automation, or agentic pipelines with real-world actuators.
Expanded weapons language: Tightened restrictions on content that could assist in the development or deployment of chemical, biological, radiological, or nuclear weapons; the language now also covers dual-use precursor guidance more explicitly.
Model wellbeing clause: A new prohibition on "sustained and needless abusive or cruel behavior toward models" is the first time a major AI company has codified protections for the model itself in a usage policy — reflecting Anthropic's ongoing model-welfare research programme.
Action required before November 12
If your product operates in civic/election contexts, previously assumed out-of-bounds by the old policy, review the new election rules — you may now have more latitude. If you are building any hardware or agentic system that takes physical actions, implement the override and fail-safe requirements before the effective date. The new policy text is at anthropic.com/legal/aup.
🧭 Claude Science Agents Build the First Complete Ultraviolet Map of the Sky — A Task Deferred for Decades
Johns Hopkins astrophysicist Brice Ménard used Anthropic's Claude Science multi-agent workbench to complete the first comprehensive ultraviolet sky map — a project that had languished for years due to its technical complexity. The result, published October 8 in an Anthropic Research post, demonstrates how a coordinated swarm of Claude agents can tackle the kind of long, multi-step scientific workflow that single-session AI interactions cannot sustain.
What the agents did
Data retrieval: Agents autonomously queried NASA's GALEX mission archives and several supplementary UV telescope catalogues, handling authentication, rate limits, and schema differences without human intervention.
Multi-source normalisation: Observations from different instruments were standardised into a common photometric system — a task that typically requires months of manual calibration work.
Artefact removal: Instrumental noise, diffraction spikes, and detector saturation were identified and flagged across tens of millions of source measurements.
Generative gap-filling: Approximately 30% of the sky had missing or corrupted observations. Claude Science applied ML inpainting — predicting flux values from surrounding regions — and achieved ~10% accuracy on hidden test areas, well within the margin useful for large-scale statistical analyses.
Why this matters for AI-assisted research
The UV sky map is the first peer-reviewable output from Claude Science's agentic research pipeline at full-sky scale. For developers building research automation or scientific data pipelines, the key insight is that the bottleneck was not model capability but workflow architecture: decomposing the task into retrievable subtasks that agents could execute independently, then synthesising the results. The techniques — dataset discovery, cross-catalogue normalisation, generative imputation — are directly transferable to genomics, climate data, and financial time-series problems.
Claude Sciencemulti-agentresearch automationUV sky mapagentic pipelinesscientific computing
🧭 Anthropic Pledges $150M Over Three Years to the White House Genesis Mission — Claude and Claude Code Go to Work for 15+ Federal Research Agencies
At the White House "Science: A New Golden Age Summit" on October 8, Anthropic announced a $150 million, three-year commitment to the Genesis Mission — a federal programme spanning more than 15 agencies (NASA, NIH, NSF, DOE, NIST, and others) that uses AI to accelerate research in fusion energy, quantum computing, drug discovery, and climate modelling. Anthropic's contribution provides Claude, Claude Code, and API credits directly to hundreds of participating research teams, alongside on-site training and technical support staff. The announcement was part of a broader White House push that totalled $2.4 billion in AI-for-science pledges from multiple technology companies.
What researchers get
API access: National lab researchers and agency scientists gain subsidised access to Claude Opus 4.6 and Sonnet 4.6 through a dedicated Genesis Mission API tier — without going through standard commercial procurement.
Claude Code: The IDE-native coding assistant will be available to software engineers at participating agencies, with a workflow designed for high-performance computing environments (SLURM jobs, container orchestration, Fortran/C++ codebases).
On-site training: Anthropic will embed technical staff at select facilities — the initial cohort includes Argonne National Laboratory and the Broad Institute — to help researchers integrate Claude into existing pipelines.
Safety guardrails: Work performed under the Genesis Mission is subject to additional export-control and dual-use review protocols beyond standard API terms.
Signal for developers
The Genesis Mission is a leading indicator of where enterprise AI usage is heading: large, long-duration research workloads where the bottleneck is not raw model capability but integration into legacy systems and workflows. If you are building tools for scientific computing audiences, this is a strong signal that government and academic institutions are ready to adopt Claude-based tooling — and that Anthropic is actively building relationships that could expand the market.
