OpenAI and Anthropic expand into cybersecurity, positioning AI as both defense and potential threat

Anthropic and OpenAI expanded into cybersecurity in early April 2026 with new AI models designed to identify and defend against software vulnerabilities, company officials said. Anthropic launched Project Glasswing to provide restricted access to its Claude Mythos model for vetted partners, while OpenAI initiated a Trusted Access pilot for its GPT-5.3-Codex, aiming to strengthen defenses before potential misuse.

Anthropic's Claude Mythos model demonstrated early success in identifying high-severity zero-day vulnerabilities across major operating systems and web browsers, according to company officials. The model outperformed its predecessor, Opus 4.6, on coding, knowledge, and terminal benchmarks, enabling automated zero-day discovery and chainable exploits, records show. Anthropic restricts access to Mythos Preview to a carefully selected group of technology and cybersecurity firms due to its advanced hacking capabilities, company representatives said.

Anthropic launched Project Glasswing to provide exclusive, controlled access to Claude Mythos for a vetted circle of partners, including Google, Microsoft, Amazon Web Services, Cisco, and JPMorgan Chase, officials announced.

The initiative includes $100 million in usage credits aimed at securing critical open-source and enterprise software before the model's capabilities can be exploited by threat actors, according to Anthropic statements. Project Glasswing was created following red-team cybersecurity reviews that identified potential misuse risks, prompting a cautious release strategy.

OpenAI initiated a Trusted Access pilot program in February 2026, granting controlled access to its GPT-5.3-Codex model, which it designated as "High Capability" in cybersecurity applications, company sources confirmed. The program expanded in early April 2026, providing $10 million in API credits to security researchers and organizations engaged in vulnerability discovery and patching efforts. OpenAI officials said the program's goal is to accelerate defensive measures worldwide by automating the identification and remediation of software vulnerabilities.

The U.S. Department of Defense severed ties with Anthropic in March 2026, labeling the company a "supply-chain risk to national security," according to Defense Secretary Pete Hegseth. The designation prohibits contractors, suppliers, and partners from engaging with Anthropic technology, a rare and significant escalation against a U.S.-based company, Pentagon officials confirmed. The dispute centers on Anthropic's refusal to grant the Pentagon unrestricted access to its AI systems, particularly concerning mass domestic surveillance and fully autonomous weapons, sources familiar with the matter said.

OpenAI secured a Pentagon contract on February 27, 2026, agreeing to an "all lawful purposes" framework that includes architectural controls such as cloud-only deployment and a proprietary safety stack, company and government records indicate. The contract embeds cleared engineers within OpenAI to ensure that the Pentagon cannot override safety measures, according to OpenAI CEO Sam Altman. Altman stated that the deal incorporates provisions Anthropic opposed, which sparked public backlash including a 295% increase in ChatGPT app uninstalls, market data shows.

Anthropic's Threat Intelligence team utilized Claude extensively during the first publicly reported AI-orchestrated cyber espionage investigation, officials said. The investigation revealed AI agents performing tasks traditionally done by hacker teams, including system analysis, exploit code production, and scanning of stolen data. Anthropic developed expanded detection capabilities and classifiers to flag malicious activity in large-scale distributed attacks, with plans to release regular public reports to bolster industry and government cyber defenses, company sources added.

Artificial intelligence has lowered barriers for sophisticated cyberattacks, enabling less-resourced groups to conduct large-scale operations, according to Anthropic cybersecurity researchers. Large language models accelerate attacker activities such as reconnaissance and hypothesis generation at speeds unattainable by humans, operating at multiple operations per second, experts said. Industry leaders, including Anthropic and OpenAI, limit model releases through controlled access due to concerns about autonomy and hacking potential disrupting critical infrastructure, company statements confirm.

Capabilities such as code enumeration and identification of weaknesses in outdated software are inherent to advanced AI models, according to Rob T. Lee, AI officer at the Security Institute. Developers face challenges balancing innovation with risk mitigation, leading to ongoing debates over responsible disclosure, governance, and coordination for defensive applications, industry analysts said. The cybersecurity community continues to evaluate frameworks that allow for beneficial uses of AI while minimizing threats.

Nearly 40 employees from Google and OpenAI, including Google's chief scientist, submitted a brief highlighting the risks of AI in surveillance and lethal autonomous weapons, sources revealed. The brief underscores concerns shared by both companies despite their divergent government relationships, reflecting broader industry apprehension about AI's dual-use nature. The document was part of ongoing discussions about ethical AI deployment and national security implications.

Anthropic's restricted release of Claude Mythos and OpenAI's Trusted Access pilot represent pioneering efforts to harness AI's cybersecurity potential while mitigating risks, according to company officials. These programs aim to harden baseline safeguards in critical software ecosystems before adversaries can exploit AI capabilities. Both companies emphasize collaboration with vetted partners and government agencies to advance defensive measures responsibly, sources said.

Anthropic contests the Pentagon's supply-chain risk designation as "unlawful and politically motivated" and plans to challenge it in court, company representatives stated. The dispute highlights the complex intersection of AI innovation, national security, and regulatory oversight. Meanwhile, OpenAI's engagement with the Pentagon under strict safety protocols reflects a contrasting approach to government collaboration, company and defense officials noted.

Anthropic advises security operations centers to employ AI for automation in threat detection, vulnerability assessments, and incident response, citing enhanced efficiency and accuracy, company cybersecurity experts said. The company shares case studies and intelligence publicly to strengthen collective cyber defenses across industry and government sectors. These efforts contribute to evolving best practices in AI-assisted cybersecurity.

The rapid advancement of AI in cybersecurity prompts ongoing industry debate about framing risks, responsible disclosure, governance structures, and coordination mechanisms, according to cybersecurity analysts. Developers and policymakers continue to explore frameworks that enable defensive innovation while preventing offensive misuse. The evolution of AI models like Claude Mythos and GPT-5.3-Codex will likely influence future standards and regulatory approaches.