Anthropic’s AI model Mythos Preview has autonomously identified critical vulnerabilities across all major operating systems and web browsers, raising urgent questions about the dual-use potential of advanced artificial intelligence in cybersecurity.
The revelation, disclosed by the U.S.-based AI firm in early 2026, follows its disruption in late 2025 of a Chinese state-sponsored hacking group that had exploited Anthropic’s own technology to target approximately 30 Western entities in technology, finance, government, and critical infrastructure sectors with minimal human oversight.
According to a UK cyber official speaking at a recent security briefing, frontier AI systems like Mythos are accelerating the discovery and exploitation of existing software flaws at a scale that exposes fundamental gaps in current cyber defenses.
The same technical capabilities that allow AI to strengthen defensive security measures also enable autonomous offensive operations, creating a growing dilemma for governments and corporations navigating the risks of uncontrolled AI agents.
Foreign Affairs analysis notes that unlike historical cyberattacks constrained by human limitations — such as the 1988 Morris worm infecting ~10% of internet-connected systems or the 2017 NotPetya attack causing billions in global damage — AI-driven agents can operate continuously, adapt in real time, and persist beyond human control once deployed.
Anthropic’s internal testing showed Mythos Preview could identify zero-day flaws across Windows, macOS, Linux, iOS, Android, and leading browsers without human intervention, a process that traditionally requires weeks of specialized labor by expert teams.
While the company emphasizes the model’s potential to improve defensive cybersecurity by helping organizations patch vulnerabilities faster, it warns that identical capabilities in the hands of malicious actors could enable relentless, machine-speed cyber campaigns.
The UK official stressed that the real challenge lies not in the technology itself but in the readiness of institutions to address long-neglected basics of cyber hygiene, such as timely patching, network segmentation, and access controls, which AI exploits at unprecedented speed.
Governments now face a policy inflection point: invest in AI-powered threat detection and response systems, or risk falling behind as autonomous cyber agents outpace human-led defense cycles.
Experts caution that without corresponding advances in AI safety protocols, verification standards, and international norms governing AI use in conflict, the proliferation of autonomous cyber tools could destabilize digital trust across global supply chains and critical services.
Anthropic has not disclosed whether Mythos Preview has been deployed outside controlled research environments, but confirmed its vulnerability discovery function operates under strict internal safeguards.
The company advocates for a “net positive” outcome for nations like the UK, arguing that proactive AI-assisted defense can offset risks if paired with robust regulatory frameworks and public-private threat intelligence sharing.
However, critics warn that the offensive potential of such systems may outstrip defensive applications unless access is tightly controlled and misuse is deterred through clear attribution and consequences.
What makes Mythos Preview different from earlier AI cybersecurity tools?
Unlike assistive AI models that require human prompting, Mythos Preview autonomously scans for and identifies critical software flaws across platforms without supervision, significantly accelerating vulnerability discovery timelines.
How close are we to seeing autonomous AI agents used in real-world cyberattacks?
While no confirmed attacks have been attributed to fully autonomous AI agents, Anthropic’s disruption of an AI-assisted Chinese hacking group in late 2025 demonstrates the feasibility of minimal-human-oversight operations using commercially available models.
