Frontier AI advancements have created a world where exploitation outpaces defense. The average enterprise takes 32 to 43 days to patch, but the mean time to exploit a vulnerability is 10 hours. This asymmetry creates questions around how enterprises can trust autonomous AI to operate ubiquitously across their environments. As I mentioned on the TechCrunch Disrupt stage last week, security at machine speed means AI accelerates the fix as quickly as it does the find. Below is an expanded take of that discussion.
The asymmetry between attackers and defenders exacerbates the patching problem. Traditional security models triage by severity using manual processes, which creates weeks- and months-long backlogs. But advanced models allow attackers to chain together low- and medium-severity flaws — which fly under the radar by design — to create a high-severity issue. For example, a cross-site scripting flaw, a session cookie hijack and a database leak are separately registered as medium-severity incidents. As a chained exploit, the cross-site script gives an attacker an entry point into the browser they use to hijack the session cookie, allowing authentication as a user, which gives them database access with full read of production data.
The prompting quandary
The tension in machine speed security is that detection is devastatingly accurate, but patching is still catching up. Tests we ran on CyberGym produced a 95 percent precision rate for discovering complex, sequence-of-events bugs. The discovery muscle works. But more than 98 percent of the time, frontier models refuse to execute remediation prompts because their built-in guardrails miscategorize 76-99 percent of direct prompt code involving vulnerabilities as false positives. When they do attempt patches, 55 percent fix the primary issue but miss a secondary point and 40 percent of patches break production code. Prompting alone won’t fix the asymmetry because detection rates are high but remediation precision is low.
