AI Has Minted a New Kind of Attacker — One Who Knows Nothing
Generative AI closes the skill gap between vague criminal intent and working malware. Responsible disclosure norms weren't built for that world.
The barrier to writing functional malware used to be meaningful. Not insurmountable, but real. You needed to understand memory, execution flow, detection heuristics. That friction kept a certain class of would-be attacker on the sidelines.
AI has removed it.
Large language models can now take vague, poorly specified malicious intent and return working code. Payload generation. Evasion logic tuned to sidestep basic security controls. Conversion of 'I want something that does X bad thing' into a deployable artifact — no programming background required. The attacker's own ignorance no longer limits them.
This creates a category that the security industry has no clean name for yet: the zero-knowledge threat actor. Not a script kiddie in the 2004 sense, who at least had to find and run someone else's tool. Something newer. Someone who couldn't describe a stack frame but can still produce a payload, because the LLM understood the intent even when the operator didn't.
The implications for vulnerability disclosure are underappreciated.
Responsible disclosure — the 90-day-and-notify model, coordinated patching windows, embargo agreements — rests on an assumption: that the window between discovery and exploitation is long enough to matter. Researchers find, vendors patch, users update. The gap is the product. It only works if that gap is larger than the time it takes a capable attacker to independently reach the same finding.
AI compresses that timeline from both ends. It accelerates attacker capability, so the pool of people who can weaponize a disclosed vulnerability grows fast. It also assists in vulnerability discovery itself, meaning the 'independent rediscovery' risk that researchers always quoted as theoretical is becoming operational.
Vendors patching on 90-day cycles weren't designed for a world where the sophistication floor dropped this sharply, this fast.
None of this is a call to abandon coordinated disclosure. Unilateral full-drop is worse. But the working assumptions baked into current frameworks — that only skilled actors exploit, that windows are measured in weeks not hours — deserve honest reexamination by the bodies that govern them, including CISA, ENISA, and the ISO/IEC JTC 1 standards track.
The threat model changed. The process hasn't caught up.
What security teams should do now
Patch aggressively and treat the 90-day window as a ceiling, not a schedule. Threat-hunt for commodity tooling, not just nation-state indicators — the attacker profile has widened. Review whether your disclosure agreements include AI-assisted exploitation risk in their severity calculations. If they don't, they're using an outdated threat model.
