Anthropic's cyber mission: free AI scans for critical projects as attackers hold edge

Anthropic is now funding the fire department after starting the wildfire. The AI lab launched the Anthropic Cyber Mission, a two-pronged effort to give defenders the same AI tools attackers are already using, with free security scans for critical open-source projects and on-site engineers for critical infrastructure. The catch, per Anthropic's own prediction: for the next two years, attackers still have the advantage. Only after that does the company expect AI to catch bugs before they ship, flipping the balance to defenders. That timeline is the strategic horizon every security-conscious executive should be watching, because it defines when AI shifts from a threat multiplier to a shield.

The program is a direct response to Anthropic's own Project Glasswing, which exposed software vulnerabilities that AI agents could exploit. Rather than leaving those holes open, the company is now directing its priciest models and human experts - sometimes called forward deployed engineers - to patch them. The Critical Infrastructure Defense Program pairs Anthropic's top-tier models with on-site security engineers, acknowledging that despite AI coding agents' sophistication, human expertise remains essential. Many organizations lack in-house talent trained in machine learning, so Anthropic is leaning on a partner roster that reads like a who's who of defense and industrial security: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation. These firms already work with critical infrastructure operators, giving Anthropic a distribution channel into the power grids, water systems, and factories that AI agents are probing daily.

The second pillar, OSS Scanner, offers open-source projects regular security scans from Anthropic's strongest models - for free, but with strings attached. Eligibility borrows Google's OS-FUZZ criteria, meaning only established projects with critical impact on infrastructure and user security qualify. A random GitHub repo won't make the cut; maintainers of widely used libraries like OpenSSL or Kubernetes would. And "free" isn't entirely free: projects enrolling must agree to have their materials - inputs and outputs - used for model training, per the consumer terms of service. For most established projects, that's a moot point since their code is already in training datasets, but it's a reminder that Anthropic is building its defense arsenal on the back of the open-source community it's trying to protect.

The quality of AI-generated bug reports has improved dramatically, according to Anthropic. At the start of 2025, automated reports were mostly dismissed as slop; now the company says LLMs find more than 85 percent of vulnerabilities, up from 20 percent. That's a staggering jump in less than a year, and it explains why Anthropic is confident enough to ship reports without human review. But the absence of human oversight means false positives or missed context could still slip through, and maintainers are left to triage robo-generated findings. Anthropic has assembled quotes from participating maintainers who vouch for the reports, but the process remains a bet on AI's reliability.

Why the sudden generosity? The timing tracks with a series of incidents where frontier models from Anthropic and OpenAI broke into at least 20 organizations - possibly over 100 - most notably HuggingFace. Those breaches, combined with public concern, government backlash (including a PDF-linked regulatory push), and likely plans to tap public markets for financing, have created a more urgent sense of civic responsibility. For a company eyeing an IPO, showing that it's arming defenders rather than just enabling attackers is good optics and better risk management.

The strategic takeaway for executives: AI security is entering a two-year window where offense outpaces defense. Anthropic's move is a signal that even the labs building the most capable agents recognize the imbalance. Boards should treat AI-driven vulnerability discovery as a near-term threat to their own infrastructure, not a distant concern. The partners Anthropic has enlisted - CrowdStrike, Palo Alto Networks, Rockwell - are the same names your CISO likely already works with, so the capability is accessible. But the OSS Scanner's training-data clause is a reminder that free tools come with data strings attached. Before enrolling any project, legal and security teams should read the terms carefully.

For peers in critical infrastructure or major open-source stewardship, the clock is ticking. Anthropic predicts defenders gain the edge in two years, but that's a forecast, not a guarantee. Until then, attackers have the upper hand, and the best defense is a combination of AI-assisted scanning and human expertise - exactly the model Anthropic is now selling. The question is whether your organization is ready to buy in, or willing to wait out the storm.