Mythos now scans code for 2 US security agencies—dual-use AI debate intensifies

When Anthropic built Mythos to find and exploit software flaws, it created one of the clearest dual-use AI tools to date. The model’s arrival at CISA, following NSA’s earlier adoption, demonstrates that even an AI lab with rigid safety commitments can become essential to national security infrastructure. That reality forces a fresh conversation about safeguards, control, and the line between defensive and offensive cyber operations in an AI-native world.

The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has begun using Anthropic’s AI model Mythos to scan government code repositories for security flaws, according to three people familiar with the matter. The deployment, carried out by CISA’s Attack Surface Evaluation team, marks a significant expansion of advanced artificial intelligence into federal cybersecurity operations — and it has already uncovered a large number of vulnerabilities, though the severity and scope remain undisclosed. The initiative underscores a growing, though still delicate, embrace of Anthropic’s technology by the U.S. government, especially given the company’s fraught history with the Pentagon earlier this year.

“

When Anthropic built Mythos to find and exploit software flaws, it created one of the clearest dual-use AI tools to date.

The background is critical. In February 2026, the Pentagon designated Anthropic as a supply-chain risk, a classification typically reserved for foreign entities suspected of espionage. The move came after Anthropic refused to remove built-in safety safeguards that prevented its AI systems from being used for autonomous weapons or domestic surveillance. The decision sparked a legal challenge, and in March, a federal judge blocked the designation, easing immediate tensions. Since then, the government’s posture has shifted markedly. The National Security Agency (NSA) has been using Mythos since at least April, and now CISA’s adoption — confirmed in July 2026 — cements Anthropic’s role in national security infrastructure.

Mythos itself is a purpose-built tool: a model designed to identify and exploit cybersecurity vulnerabilities. Its deployment inside CISA’s code audit pipelines represents a dual-use frontier. On one hand, the model can dramatically accelerate the discovery of flaws that foreign intelligence agencies or cybercriminals might exploit; on the other, the same capability, if misused or misaligned, raises profound safety questions. Anthropic’s refusal to remove its safeguards — and the subsequent reversal of the Pentagon blacklist — indicate that the government may now accept, or at least tolerate, AI models with built-in ethical constraints. That is a noteworthy evolution from earlier fights over AI-liberties versus national security.

The implications for federal cybersecurity are substantial. CISA’s Attack Surface Evaluation team conducts digital security assessments and hacking exercises across the entire U.S. government. Integrating Mythos means these assessments can be done at machine speed, with the ability to comb through millions of lines of code in search of subtle vulnerabilities. The fact that two sources said the audits have already uncovered a large number of flaws, even without specifics, suggests the technology is delivering real, actionable findings. This could force a broader rethink of how the government approaches software supply chain security, already a top priority after high-profile breaches.

For Anthropic, the timing aligns with its confidential IPO filing. Government contracts with high-profile agencies like the NSA and CISA validate the company’s technology in a mission-critical context, likely boosting investor confidence. However, the supply-chain risk episode remains a cautionary tale. Political winds can shift, and a future administration or different defense leadership could resurrect the blacklist if the company refuses to bend on safety guardrails. The dual-use nature of Mythos — poised between offensive cybersecurity tool and a potential weapon itself — keeps Anthropic in a perpetual tension between its safety-first brand and the demands of national security clients.