32 lawmakers told the White House this week: AI is now finding bugs faster than companies can patch them.

WHAT THE LETTER ACTUALLY SAYS

The May 13 letter from 32 House lawmakers to National Cyber Director Sean Cairncross makes a direct argument. AI models like Anthropic's Mythos and OpenAI's GPT-5.5-Cyber are demonstrating capability to identify software vulnerabilities at scale and at speed. The capability exists. The capability is in the hands of both defenders and adversaries. The federal government's traditional cyber coordination playbook was not designed for what happens when one of these models surfaces thousands of unknown vulnerabilities in a coordinated disclosure, or when adversaries use the same capability to develop exploits faster than patches can ship.

The letter outlines seven recommendations. Coordinate federal responses to large volumes of AI-generated vulnerability disclosures. Help critical infrastructure operators deploy patches faster than current cycles allow. Expand controlled defensive access to frontier cyber models through a voluntary framework. Clarify the standards for how advanced cyber models are shared, restricted, and accessed across government and industry. Coordinate with Treasury and the National Economic Council on access decisions. Identify areas where Congress may need to legislate. Address the risk that adversaries could steal, copy, or replicate comparable systems.

Rep. Bob Latta (R-Ohio) led the effort. The signatories cross party lines. The letter was first shared with Axios on May 13 and reported across Nextgov, CyberScoop, and MLex through the day. The framing is sober. The lawmakers explicitly write that capabilities have outgrown the traditional cyber coordination playbooks.

THE MYTHOS THREAD

The letter cites Anthropic's Mythos by name. The reference matters because it threads through a much longer story than this week's letter alone. Anthropic delayed the Mythos rollout in April due to cybersecurity concerns, specifically the model's ability to uncover and weaponize software vulnerabilities at scale. Anthropic limited access through a controlled program called Project Glasswing. The White House opposed an Anthropic plan to expand access. The Pentagon is in active litigation with Anthropic over whether the Defense Department can declare the company a supply-chain risk. Defense Secretary Hegseth and the Trump administration outlined a six-month phase-out for government use of Anthropic's tools earlier this spring.

The Atarus blog covered the Mythos breach on April 24, when an unauthorized Discord group was found to have accessed the model ahead of its public release. The April 24 post argued that Mythos was the inflection point for AI-augmented offensive operations. The lawmakers' May 13 letter is the political confirmation of that argument. The capability the Atarus body of work named in April is the capability Congress is now citing as the reason for federal coordination.

The letter's framing is intentionally bipartisan. Senate Minority Leader Chuck Schumer wrote separately to DHS Secretary Mullin earlier this week urging closer coordination with state and local governments on AI cyber threats. Schumer explicitly said AI systems will soon be better than humans at finding software vulnerabilities and that critical infrastructure operators need federal preparation. The House Homeland Security Committee received a closed-door briefing this week led by Anthropic's frontier red team and national security programs team, with Mythos demonstrated directly to lawmakers. The political layer is moving on multiple tracks simultaneously.

WHY THIS WEEK IS THE INFLECTION

The Atarus blog has now documented thirteen significant security incidents in eighteen days. The arc has been technical. The MCP architectural RCE on April 21. The Mythos breach on April 24. LMDeploy SSRF on April 25. GlassWorm sleeper extensions on April 26. Itron utility breach on April 27. Entra Agent ID privilege escalation on April 28. LiteLLM SQL injection on April 29. PocketOS database wipe on April 30. PyTorch Lightning supply chain on May 1. Bleeding Llama Ollama memory leak on May 6. Dragos Monterrey water utility on May 7. GTIG AI-developed zero-day on May 12. TeamPCP Checkmarx Jenkins plugin compromise on May 13.

Each incident was a different layer of the stack. AI tooling protocols, inference servers, vendor identity, marketplaces, package supply chains, coding agents, critical infrastructure intrusions, AI-developed exploits, supply chain attacks against security vendors. The Atarus blast radius synthesis on May 10 argued the structural pattern. The eleven incidents at that point were all deployment architecture failures rather than capability failures. The capability was within the design envelope in every case. The blast radius determined severity.

The federal response is now catching up. CISA released CI Fortify on May 5, which is the operational guidance framework for critical infrastructure operators to plan for the threat environment the Atarus body of work has been documenting. The May 13 lawmaker letter is the legislative branch's escalation of the same concern. The Schumer letter to DHS is parallel pressure from the Senate. The Homeland Security Committee Mythos briefing is the third track. The executive branch is weighing executive action on AI cybersecurity. The political layer is now actively responding to the technical reality.

For practitioners, the implication is that regulatory action on AI cybersecurity is approaching. The timing is uncertain. The direction is not. Organizations that begin building the operational disciplines now (blast radius reviews, AI-augmented code review for semantic logic flaws, vendor risk management that accounts for sustained compromise campaigns, CI Fortify-compatible isolation and recovery planning) will be positioned ahead of whatever regulatory framework eventually emerges. Organizations that wait will absorb the regulatory action reactively.

WHAT THE LAWMAKERS ACTUALLY WANT

The seven recommendations in the May 13 letter map directly to the operational gaps the past three weeks of incidents have surfaced. Each one has practitioner translation.

First, coordinate federal responses to large volumes of AI-generated vulnerability disclosures. The expected scenario is that an AI model coordinated by a security firm, a research lab, or a federal program identifies thousands of vulnerabilities across a large software ecosystem and discloses them through a coordinated process. The federal coordination question is who triages, who notifies vendors, who deploys patches, who tracks adoption, and who communicates with critical infrastructure operators when their software is affected. The current coordination mechanism (CISA Vulnerability Disclosure Policy, CVE Numbering Authorities, vendor security advisories) was built for a steady-state flow of dozens to hundreds of disclosures per month. The threat environment lawmakers are describing produces an order of magnitude more volume.

Second, help critical infrastructure operators deploy patches faster than current cycles allow. This is the operational concern Schumer's parallel letter to DHS named explicitly. Water utilities, transportation, energy, healthcare, and other critical infrastructure operators run on patch cycles measured in quarters rather than days. AI-augmented attackers operate on exploit development cycles measured in hours. The asymmetry is structural, not a problem of insufficient effort. Federal support for accelerated patching at critical infrastructure operators is the practitioner ask the letter implicitly endorses.

Third, expand controlled defensive access to frontier cyber models. Project Glasswing, OpenAI's Trusted Access for Cyber program, and comparable frameworks at other AI providers currently restrict access to a small set of vetted users. The lawmakers want the framework expanded so that critical infrastructure operators, federal agencies, and qualifying private firms can use these models defensively. The expansion is voluntary in the letter's framing. The implication is that mandatory frameworks may follow if voluntary expansion proves insufficient.

Fourth, clarify standards for how advanced cyber models are shared, restricted, and accessed. The current standard is ad hoc. Each provider makes its own access decisions. The federal government engages with each provider separately. The lawmakers want a coherent framework that applies consistently. This is the closest the letter comes to advocating for regulation, though framed as standards clarification rather than regulatory authority.

Fifth, coordinate with Treasury and the National Economic Council on access decisions. The recommendation acknowledges that access decisions about frontier AI models are not purely cybersecurity questions. They are also competitiveness, trade, and national economic policy questions. The Treasury and NEC coordination ask reflects an understanding that AI cyber capabilities have implications for U.S. strategic competition with China specifically.

Sixth, identify areas where Congress may need new legislation. The letter is an executive branch ask, not a legislative declaration. The acknowledgment that legislation may be needed reflects awareness that executive action alone may be insufficient for the framework the lawmakers want.

Seventh, address the risk that adversaries could steal, copy, or replicate comparable systems. This is the security-of-the-models question. If Anthropic's Mythos is held back from broad access but is also a target for theft and replication, the federal response needs to include defensive protection of the labs themselves. The Atarus blog covered the operational version of this concern on April 24 with the Mythos breach. The lawmakers are naming the structural version.

INDICATORS AND TIMELINE

WHAT THIS MEANS FOR DEFENDERS