AI Security Testing Enters a New Phase
AI security testing is rapidly moving from research labs into frontline cybersecurity workflows. Vendors are racing to build vulnerability detection AI that can comb through sprawling codebases, uncover subtle flaws and propose fixes long before human teams would normally find them. Anthropic’s Mythos model and Google’s CodeMender agent exemplify this new generation of automated security tools, designed specifically for cybersecurity testing automation rather than general-purpose coding assistance. The shift matters because modern enterprise stacks—especially complex operating systems and cloud platforms—produce too much code and configuration drift for manual review alone. AI promises to continuously scan, reason about and stress-test software, turning what used to be sporadic penetration tests into ongoing analysis. Yet this acceleration raises uncomfortable questions: how do you prevent the same systems that strengthen defense from being repurposed for offense, and where should human oversight sit in an increasingly automated discovery pipeline?
Mythos and the Mac: A Benchmark for Automated Discovery
Anthropic’s Mythos AI has set a new benchmark for AI security testing by helping researchers uncover critical macOS vulnerabilities. Security firm Calif used an early Claude Mythos Preview to identify a sophisticated exploit chain that targets memory in Apple’s desktop software, linking two distinct bugs to achieve a privilege escalation exploit. This kind of chained attack, which bypasses standard protections to reach restricted parts of the operating system, has historically required elite human expertise and extensive manual testing. Mythos’ role highlights how vulnerability detection AI can navigate deeply layered, closed platforms that have been harder to probe at scale. Anthropic’s response underscores the dual-use risk: the company has warned that Mythos is so effective at finding flaws that broad release could threaten digital infrastructure. Instead, it operates Mythos under Project Glasswing, a controlled-access program that shares capabilities with select partners for defensive purposes while keeping exploit discovery and patching under strict human review.
Google’s CodeMender: Expanding Access Without Going Fully Public
Google’s CodeMender agent is emerging as a direct competitor to Mythos in automated security tools, but with a distinct focus on integration into real engineering pipelines. Developed by Google DeepMind, CodeMender uses Gemini Deep Think models alongside static and dynamic analysis, differential testing, fuzzing and SMT solvers to trace vulnerabilities to their root causes. It then drafts patches and tests them before any human ever sees the proposed fix. Google is widening API access for vetted security experts, allowing them to embed CodeMender into existing triage, validation and release workflows. This controlled rollout keeps the tool out of general release while still broadening practical experience with AI-powered cybersecurity testing automation. Crucially, every patch remains subject to human review and must pass repository rules, regression checks and internal change controls. Google’s cautious expansion mirrors Anthropic’s gated Mythos and Claude Code Security previews, signaling that access policy is now a core competitive dimension alongside raw model capability.
Redrawing the Vulnerability Discovery Landscape
The rise of Mythos and CodeMender is reshaping how enterprises think about vulnerability discovery. Instead of relying primarily on periodic human-led audits, organizations can deploy AI security testing agents that continuously analyse code and runtime behaviour. On complex platforms like macOS, where proprietary architectures and intricate memory management once limited automated coverage, Mythos’ exploit-chain discovery shows that AI can now surface issues that might elude even seasoned researchers. In parallel, CodeMender demonstrates how vulnerability detection AI can be woven directly into software delivery pipelines, not just used in isolated security exercises. By combining program analysis with generative patching, it moves the industry closer to closed-loop detection and remediation. Yet both systems remain deliberately constrained: access is limited, usage is monitored, and outputs are channelled through defensive teams. As more vendors adopt similar patterns, the competitive race is less about unleashing unrestricted tools and more about who can safely accelerate automated discovery for trusted enterprise users.
Why Human Oversight Still Anchors AI-Driven Security
Despite the speed and scale promised by cybersecurity testing automation, Anthropic and Google are aligned on one principle: human oversight remains non-negotiable. Mythos operates under Project Glasswing, where select partners receive detailed reports—such as the 55-page disclosure delivered to Apple—but human security engineers decide how to prioritise, validate and patch the findings. Anthropic’s Claude Code Security similarly positions AI as a patch-suggestion system, leaving final approval to expert reviewers. CodeMender follows the same pattern. Even though it can propose and test fixes, human maintainers retain the final say on whether changes ship to production. This oversight is not just about safety; it reflects practical realities of software governance, where policy, risk appetite and operational constraints shape every release. As vulnerability detection AI matures, the most effective organisations will likely treat it as an expert co-pilot—relentless in discovery and draft remediation, but always subject to human judgment before any security-critical change is deployed.