Two of the leading artificial intelligence companies have thrust the cybersecurity world into urgent recalibration. Anthropic unveiled Project Glasswing in early April, restricting its powerful Claude Mythos Preview model to a select group of partners. Days later OpenAI responded with its own tiered access program for GPT-5.5 and a dedicated cybersecurity platform called Daybreak. The moves mark a sharp turn. No longer content to supply tools that write code, the AI labs now position themselves as central players in the defense against the very threats their technology amplifies.
Claude Mythos Preview changed the conversation almost overnight. The unreleased frontier model identified thousands of high-severity vulnerabilities, many of them zero-days. Some had survived decades of human scrutiny and millions of automated tests. It found a 27-year-old flaw in OpenBSD that allowed remote crashes. It uncovered a 16-year-old bug in FFmpeg missed by tools that hit the code five million times. The model even chained vulnerabilities in the Linux kernel to escalate from user access to full machine control, all autonomously. “Claude Mythos Preview is a general-purpose, unreleased frontier model that reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities,” Anthropic stated in its announcement. The company committed up to $100 million in usage credits and $4 million in donations to open-source security efforts through the Linux Foundation and Apache Software Foundation.
Partners in Project Glasswing read like a who’s who of technology infrastructure. Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, NVIDIA, Palo Alto Networks and the Linux Foundation signed on. Over 40 additional organizations maintaining critical software gained access too. The goal remains explicit. Give defenders a head start before the same capabilities spread to adversaries. Anthropic has no plans for general release. Instead it funnels the model through monitored API channels on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.
OpenAI took a different path. Its Trusted Access for Cyber program scales access more broadly to verified defenders. GPT-5.5-Cyber supports specialized workflows while stricter tiers handle general enterprise work. Then came Daybreak. The platform embeds GPT-5.5 and Codex Security directly into development and security pipelines. It analyzes codebases, models attack paths, validates exploits in isolated environments, and proposes patches. Integrations span Cisco, Cloudflare, CrowdStrike, Palo Alto Networks, Oracle, Fortinet, Zscaler, Akamai and more. Three model variants address different risk levels. Standard for routine tasks. Trusted Access for defensive security. And a controlled permissive version for red teaming and penetration testing.
The rivalry exposes deeper disagreement on deployment strategy. Anthropic chose tight limits and a small circle of trusted organizations, warning that the model presages a wave of systems that exploit flaws faster than humans can respond. OpenAI pursues wider rollout under verification, monitoring and human oversight. Both approaches reflect the same pressure. Attackers already operate at machine speed. Mandiant’s M-Trends 2026 report showed the window from vulnerability disclosure to exploitation collapsing. In some cases it turned negative. Attackers strike before patches exist. Human teams cannot keep pace.
Concerns extend beyond the models themselves. AI coding assistants from both companies accelerate software creation at unprecedented rates. Developers generate millions of lines of new code, often laced with subtle errors. The vulnerability surface expands rapidly, especially as organizations pull in open-source libraries. One flaw can cascade across thousands of downstream users. Isaac Evans, CEO of Semgrep, captured the anxiety. “Everyone’s predicting that there will be a lot more hacking this year.” Feross Aboukhadijeh, CEO of Socket, added that the vulnerability surface “is expanding really quickly.”
Chief information security officers feel the strain. Manoj Nair, who leads emerging technologies at Snyk, described the environment as “AI fog.” Teams scramble to incorporate new tools while scanning legacy codebases flooded with fresh weaknesses. Yet partnerships form quickly. Security vendors that once competed now integrate OpenAI and Anthropic capabilities. Startups such as Semgrep, Socket and Snyk work alongside the model providers. Mozilla used Mythos to fix more bugs in a single period than in the prior year combined.
Government officials took notice. OpenAI and Anthropic each delivered separate briefings to staff of the House Homeland Security Committee in late April. The closed-door sessions covered the models’ implications for critical infrastructure, particularly sectors with limited resources. House Homeland Security Chair Andrew Garbarino called the engagements “productive partnerships between industry and government.” Other lawmakers expressed alarm. One described the demonstration as “very scary” and stressed the need for guardrails. Another noted how readily accessible the technology appears, raising fears that the wrong actors could obtain it.
The European Union entered the discussion as well. OpenAI offered the bloc preview access to its cybersecurity model. Anthropic has held back so far, though talks continue. The contrast highlights differing philosophies on sharing powerful capabilities with public-sector reviewers. Meanwhile the Pentagon has navigated its own tensions, at one point dropping Anthropic over security concerns before engaging OpenAI under stricter controls.
Industry experts watch the competition with mixed reactions. Mitch Ashley, vice president at The Futurum Group, observed that Daybreak positions OpenAI as a control surface for application security. Doug Merritt, CEO of Aviatrix, welcomed the compression of discovery-to-patch cycles from days to minutes but cautioned that architecture problems often matter more than individual patches. Logan Graham, who heads Anthropic’s frontier red team, emphasized collaboration. “Security is always a team sport.”
What emerges is an arms race defined by speed and control. The same agentic reasoning that lets models hunt vulnerabilities also lets them generate exploits without constant human direction. Defenders hope to stay ahead by embedding these systems into their workflows first. Yet the technology diffuses. Benchmarks show both Mythos and GPT-5.5 achieving strong results on cyber tasks. The UK government urged boards to treat cyber risk as a standing agenda item rather than something delegated to IT.
OpenAI’s Daybreak launch, reported across outlets including DevOps.com, signals deeper enterprise integration. The platform does not simply scan. It models threats, tests in sandboxed settings, and feeds results back into development loops. Security vendors embed it. That suggests the AI layer is becoming foundational rather than optional. Similar dynamics play out in Anthropic’s consortium. Partners share learnings. Public reports on fixed vulnerabilities are promised within 90 days. The effort aims to harden foundational systems that run banks, hospitals, power grids and logistics networks.
Questions linger about long-term balance. If models improve at similar rates, does one company’s more restrictive stance provide real advantage? Or does broader distribution under governance allow faster collective defense? Both labs continue to brief lawmakers and agencies. Both invest heavily in safeguards that detect and block dangerous outputs. Progress on those safeguards will determine how widely these systems can eventually deploy.
The events of spring 2026 have clarified one reality. Cybersecurity no longer operates primarily at human scale. AI systems now probe codebases, chain exploits and suggest remediations faster than teams can review them. Companies that treat these models as external tools risk falling behind those that weave them into core operations. The frenzy, as Business Insider described it, shows no sign of slowing. New vulnerabilities surface daily. The models that find them also help close them. Who gains the upper hand may depend less on who builds the smartest system and more on who deploys it most effectively across the defensive ecosystem.
And the clock keeps ticking. Twenty-two seconds. That is sometimes all the gap that remains between discovery and exploitation. In that narrow window the difference between defender and attacker grows thinner. The AI companies have stepped into the breach. Whether their intervention tilts the field toward safety or simply accelerates the contest remains the central uncertainty facing security leaders today.