AI Security Testing Moves from Concept to Competitive Battleground
AI security testing is shifting from experimental labs into a fierce competitive arena for major AI companies. Instead of relying solely on human penetration testers and manual code reviews, organisations are beginning to deploy vulnerability detection AI that can sift through complex software stacks at machine speed. The catalyst is a new breed of security-focused models, led by Anthropic’s Mythos and Google’s CodeMender, which are explicitly designed to spot and help exploit software weaknesses. Both tools are guarded behind tight access controls, reflecting a shared concern: the same systems that can harden critical infrastructure could also be weaponised if released openly. This tension is driving a cautious, invite-only approach even as demand from enterprises grows. As tech giants race to prove their models can find more automated security flaws, AI is rapidly moving from assistant to primary engine in vulnerability discovery workflows.
Anthropic Mythos Security: Outsmarting macOS and Apple’s Defenses
Anthropic Mythos has quickly become a benchmark for what AI security testing can achieve. In a tightly controlled preview programme, security firm Calif used an early Claude Mythos model to uncover a sophisticated exploit chain in Apple’s macOS kernel on M5 hardware. Mythos helped researchers locate two related bugs, then assisted in building a memory corruption exploit that chained them together to bypass standard protections and achieve local privilege escalation—granting full access to areas of the system that should remain off-limits. The attack effectively outsmarted Apple’s existing security systems in a way researchers say had not been seen before. Yet Mythos did not operate in isolation: experts emphasise that human exploit developers were essential to steering, validating, and safely disclosing the findings. Anthropic’s response has been to keep Mythos restricted, arguing that its unprecedented skill at finding security flaws demands strong safeguards.
Inside Mythos: How an AI Becomes a Vulnerability Hunter
What makes Anthropic Mythos so potent for vulnerability detection AI is its ability to learn classes of bugs and generalise. According to Calif’s account, once Mythos understood how a certain category of memory bugs tended to appear in system code, it could quickly scan for similar patterns elsewhere and suggest promising attack surfaces. Rather than blindly fuzzing, Mythos contributed targeted hypotheses about where macOS might mishandle memory, guiding researchers toward the kernel issues that later became a full exploit. This reflects a broader shift: models are evolving from generic code assistants into specialised security analysts capable of reasoning about exploit chains. Still, Mythos is not an autonomous hacker. Human teams design the testing strategy, evaluate AI suggestions, and craft proof-of-concept attacks. Anthropic channels these capabilities into initiatives like Project Glasswing, positioning Mythos as a tool for defensive research while maintaining strict oversight around access and use.
Google CodeMender: Gemini-Powered Counter to Anthropic’s Lead
Google is responding with CodeMender, an AI security agent from Google DeepMind that directly targets the same emerging niche as Anthropic Mythos. Initially unveiled in October 2025, CodeMender is now entering a wider but still restricted testing phase, granting selected expert teams API access without a full public launch. The system combines Gemini Deep Think models with static and dynamic analysis, fuzzing, differential testing, SMT solvers, and other program-analysis tools. Its goal is to trace vulnerabilities back to their root causes, then draft patches that can be automatically tested before a human reviews and approves them. Google frames CodeMender as a way to help secure the world’s codebases, but, like Anthropic, it is acutely aware of dual-use risks. By expanding access cautiously, Google is both scaling vulnerability detection AI for real-world use and positioning CodeMender as a direct competitor to Anthropic’s Mythos previews.
Automation with Oversight: What This Arms Race Means for Enterprises
Despite their rivalry, Anthropic and Google share a design philosophy: use automation for speed and coverage, but keep humans in charge of judgment. Mythos is deployed in tightly controlled programmes where expert researchers guide its analysis and validate every exploit path before reporting to vendors like Apple. CodeMender similarly drafts fixes, but every patch passes through human reviewers before it can be merged, ensuring developers retain control over security decisions. For enterprises, this signals a future where AI security testing becomes a standard part of the software lifecycle, running continuously to uncover automated security flaws long before they’re discovered in the wild. As organisations grapple with sprawling codebases and evolving threats, the competition between Anthropic Mythos security tools and Google CodeMender suggests accelerating investment in AI-driven vulnerability detection—augmented, not replaced, by human expertise.