AI-assisted vulnerability discovery
In plain wordsUsing AI to find security holes in software — and often to suggest fixes — faster than human experts alone, so that defenders, or attackers, can act on them.
The use of machine learning, especially language-modellarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. agentsAI agent, n.: An AI system that does not just answer but acts: given a goal, it decides on steps, uses tools such as a browser or code editor, checks the results and keeps going until the job is done., to find exploitable flaws in software and often to generate and test patches. Systems combine model reasoning over source code with classical program analysis: they read code to hypothesize bugs, write inputs or test harnesses to trigger them (including steering fuzzers), run the program under memory-error detectors to confirm crashes, and propose fixes checked by tests. The capability is dual-usedual-use, adj.: Describes technology or knowledge that can be used both to help and to harm: the same AI that finds software bugs for defenders can find them for attackers.: what helps maintainers fix flaws before release can help attackers find exploitable ones, so frontier developers evaluate it, and the volume of AI-found bugs strains disclosure and patching capacity. Milestones include Google's Big Sleep agent finding an exploitable memory-safety bug in the SQLite database in 2024, and DARPA's AI Cyber Challenge (AIxCC), whose finalists' systems in 2025 found 54 of the vulnerabilities planted in open-source test projects, patched 43, and also found 18 real ones.
AI-assisted vulnerability discovery flagged the parser bug, wrote a crashing input and drafted a patch for human review.
Origin
Machine learning has guided fuzzing and static analysis since the 2010s, and DARPA's 2016 Cyber Grand Challenge pitted automated systems against each other. Language-model agents arrived in 2024, when Google Project Zero and DeepMind reported Big Sleep's first real-world find in November. DARPA's AI Cyber Challenge concluded in August 2025; Team Atlanta won, and the finalists' systems were open-sourced.
Related
- AI-enabled cyberattackAI-enabled cyberattack, n.: A computer attack in which criminals or spies use AI to help — writing convincing scam messages, finding weak spots, producing malicious code, or even running much of the hacking by itself.
- dual-usedual-use, adj.: Describes technology or knowledge that can be used both to help and to harm: the same AI that finds software bugs for defenders can find them for attackers.
- coding agentcoding agent, n.: An AI agent that writes and fixes software by itself: it reads a codebase, edits files, runs commands and tests, and keeps working until a feature is built or a bug is fixed.
- CybenchCybench, prop. n.: A test of AI agents' hacking skills built from 40 professional 'capture the flag' security puzzles, used to judge how much an AI could help cyberattacks or cyber defense.
- AI supply chain securityAI supply chain security, n.: Protecting everything an AI system is built from — downloaded models, datasets, code libraries and tools — against tampering, so a poisoned or fake part cannot sneak into the finished product.
- dangerous capability evaluationdangerous capability evaluation, n.: A test of whether an AI system can do things that could cause severe harm in the wrong hands or on its own — such as helping make a bioweapon, hacking computers or copying itself.
Sources
- DARPA (2025). AI Cyber Challenge marks pivotal inflection point for cyber defense. DARPA News, August 2025. darpa.mil/news/2025/aixcc-results
Mentioned in
AI securityAI security, n.: Protecting AI systems — their data, models and surrounding software — from being attacked, stolen or tricked, and dealing with the ways AI itself can be used to attack others., AI-enabled cyberattackAI-enabled cyberattack, n.: A computer attack in which criminals or spies use AI to help — writing convincing scam messages, finding weak spots, producing malicious code, or even running much of the hacking by itself., dual-usedual-use, adj.: Describes technology or knowledge that can be used both to help and to harm: the same AI that finds software bugs for defenders can find them for attackers., CybenchCybench, prop. n.: A test of AI agents' hacking skills built from 40 professional 'capture the flag' security puzzles, used to judge how much an AI could help cyberattacks or cyber defense., ExploitGymExploitGym, prop. n.: A test that measures whether AI agents can turn a known software security flaw into a working attack, using hundreds of real vulnerabilities., export control directiveexport control directive, n.: A US government order, used against AI models in June 2026, telling a company to stop foreigners from using some of its AI models for national-security reasons., Project GlasswingProject Glasswing, prop. n.: An Anthropic program, launched in 2026, that gave a hacking-capable AI model only to organizations defending important software instead of releasing it to everyone., vibe hackingvibe hacking, n.: Using AI coding assistants to carry out cyberattacks by describing the goal in everyday language, letting the AI do technical work the attacker could not do alone.