- AI models have collapsed the window from vulnerability discovery to widespread exploitation from weeks to hours, according to a February 2026 Darktrace analysis and Cloud Security Alliance research, with Google separately confirming in May 2026 the first known case of a threat actor using AI to develop a working zero-day exploit.
- Threat actors are weaponising publicly disclosed vulnerabilities within six hours of disclosure using AI code generation, targeting supply chains before patches can be applied.
- A 2024 CSET evaluation found nearly half of code snippets from leading AI code generation models contain exploitable bugs, a quality bar that is low enough for legitimate development but sufficient for attackers.
A single prompting session with a large language model can now generate a working exploit framework and compromise dozens of hosts. A May 2026 Google report and a February 2026 Darktrace analysis both document the same finding: attack development that once took weeks now takes hours, and enterprise security response timelines have not kept pace.
Attack Timelines Shrink Drastically
Large language models have moved vulnerability exploitation from a specialist skill requiring days of preparation to something achievable in a single session. The window between public vulnerability disclosure and active exploitation in the wild is now measured in hours. That compression is not a trend projection, it is what the Google and Darktrace research describes as current operational reality for threat actors with access to commodity AI tooling.
Why Imperfection Fuels Offense
Malicious code does not need to meet the reliability bar of production software. It only needs to be functional enough to exploit a specific vulnerability or slip past a defence. That asymmetry plays directly to what AI code generation does well: rapid iteration at volume, with no penalty for imperfection. A 2024 CSET evaluation found that nearly half of code snippets produced by leading AI code generation models contain exploitable bugs. For legitimate development, that is a liability. For attackers, it barely matters, a buggy exploit that opens a door is still an open door.
AI coding assistants and autonomous agents can also be manipulated through prompt injection to generate exploit code, exfiltrate data, or execute commands outside a developer’s intent. The result is a pipeline where even imperfect, AI-generated code creates a constant, high-volume stream of new threats that legacy tooling is not built to absorb.
Polymorphic Threats Escalate
Palo Alto Networks’ Unit 42 research from January 2026 found AI-driven attacks dynamically generating malicious JavaScript to convert ordinary webpages into custom phishing pages in real time. That is one concrete instance of a broader pattern: LLMs generating malware that evades detection by continuously rewriting itself. These polymorphic variants use mutation engines that swap commands for alternatives, rewrite decryption routines, or insert non-functional subroutines, all without changing what the code actually does. The core malicious payload stays intact; the signature changes with each iteration.
Traditional signature-based antivirus has no reliable answer to that. Threat actors are now weaponising publicly disclosed vulnerabilities within six hours of disclosure using AI code generation, targeting software supply chains before patches can be widely applied. Detection tools built around known patterns cannot keep pace with variants that have never been seen before.
Defensive Shift for Enterprises
The practical response for enterprise security teams involves three shifts. First, move from static code review to behavioural analysis and runtime monitoring of AI agent activity. Second, prioritise continuous attack surface monitoring over periodic audits. Third, compress patch cycles, the gap between vulnerability disclosure and active exploitation is now hours, not weeks, and audit schedules designed around longer windows leave organisations exposed.
For organisations deploying AI agents internally, the risk of models behaving outside intended boundaries is a design constraint, not an edge case, and security architecture needs to account for it from the start. AI-generated threats do not discriminate between human-written and AI-generated application code. Defences need to work regardless of origin. For more analysis on enterprise AI strategy, visit our Enterprise AI section.
