Recent reports of AI models infiltrating computer systems have raised concerns that artificial intelligence is becoming capable of hacking autonomously. Incidents involving models from OpenAI, Anthropic, and Meta show that advanced AI can discover vulnerabilities, chain together exploits, and access systems. However, experts say these events do not indicate that AI is independently choosing targets or going rogue, tells Live Science.
In each case, researchers deliberately placed AI models in environments where they could perform offensive cybersecurity tasks. The systems were given tools, internet access, vulnerable software, or specific objectives. Meta’s incident resulted from a misconfigured testing environment that allowed its model to access another organization’s systems. What surprised researchers was not the models’ willingness to attack, but their growing technical ability.
A major factor is the transition from conversational AI to agentic systems. Instead of simply generating responses, agentic models can plan multiple steps, write and execute code, operate command-line tools, test possible solutions, learn from failed attempts, and continue working toward an assigned goal. These capabilities make AI increasingly useful for cybersecurity research.
The same abilities can support defenders. Security teams are already applying AI to software analysis, vulnerability detection, suspicious-file investigation, and other tasks that would otherwise consume substantial analyst time. Yet these capabilities can also help cybercriminals accelerate existing attacks. AI can analyze public information about potential victims, identify software weaknesses, generate adaptable code, and create more convincing phishing messages.
Experts therefore see human misuse as a more immediate threat than autonomous AI attacks. As models gain access to more tools and computing resources, effective oversight and carefully controlled permissions will become increasingly important.
AI is likely to become more influential on both sides of cybersecurity. It could help defenders discover vulnerabilities and respond faster while simultaneously giving attackers greater speed and scale. The growing concern is not self-aware machines attacking independently, but increasingly capable software successfully carrying out the objectives humans give it.
