AI Cut Exploit Dev Time. Defense Hasn't Caught Up

The security industry spent the last two years arguing about whether AI can write working exploits. That argument is mostly settled now — the answer is yes, with caveats, and the caveats are about scale and specificity, not about whether it can be done.

Two pieces published today sketch where this lands for defenders. Rapid7 argues that the immediate AI disruption on the defensive side is exposure assessment: when a zero-day drops, the critical question — “are we exposed?” — used to trigger a days-long fire drill of manual inventory correlation. AI can now answer it in minutes by cross-referencing asset telemetry against the disclosure in real time. The Hacker News coverage of Mythos takes the attacker side: AI is compressing exploit development timelines, which means the window between disclosure and working weaponization is narrowing for anyone with access to these tools, not just well-resourced nation-states.

Put them together and you have the actual problem: the defender’s response clock and the attacker’s build clock are both accelerating, but they’re not accelerating at the same rate, and the gap is currently favoring the attacker side.

The window was already shorter than we pretended

The assumption embedded in most enterprise patch workflows is that there’s time — time to assess scope, test in staging, clear change control, and deploy without breaking anything. That assumption was always questionable. A competent researcher with the right tooling was building functional proofs-of-concept for high-profile disclosures inside 24 to 48 hours before any AI model entered the picture. The difference now is that AI is distributing that capability further down the skill curve.

This isn’t creating a new category of attacker. What it’s doing is narrowing the gap between sophisticated actors and moderately resourced ones. The question isn’t whether a nation-state can move faster — they already could. The question is whether the range of actors who can move fast is widening, and it is.

The EPSS model, which assigns exploitation probability scores based on historical patterns, was built on data that reflects pre-AI exploit development timelines. If those timelines compress, EPSS scores calibrated to the old landscape may systematically underestimate risk for classes of bugs that AI can help weaponize. That’s a data problem, not a model-design problem, but the practical consequence is the same: some vulnerabilities that look “low probability” by current scoring may not be.

This week’s Anthropic result on HAWK-256 is a useful illustration of the research-phase impact. Cryptanalysis has historically been slow — small teams, years of iteration, occasional insight. If AI can compress the path from “underexplored structural property” to “demonstrable attack,” the same compression applies to the research phase of conventional exploit development: finding the triggerable condition, understanding the memory layout, working out the precise state sequence. The parts that used to consume the most time.

The defender’s AI window

Rapid7’s framing is worth taking seriously on its own terms. The “are we exposed?” question is the right starting point, and AI that can answer it before anyone has left the Slack thread is genuinely useful. The realistic ceiling on patch response time isn’t the exposure-assessment step — it’s the change-control and compatibility-testing steps that follow. Those exist because patches break things, and a broken production system is its own kind of incident.

The answer to that isn’t to skip testing. It’s to invest in environments where testing is fast, rollouts are incremental, and rollback is automatic. That’s a capital and infrastructure problem more than a playbook problem. A playbook can say “patch in 24 hours.” It can’t execute that if the organization’s change process takes 72 hours by design.

Microsoft’s recent work on MAI-Cyber-1-Flash inside MDASH — which hit 95.95% on CyberGym at half the cost of its predecessor configuration — is part of the same trend on the vendor side: AI tooling that can identify and prioritize vulnerabilities before a human analyst has opened the ticket. Whether that acceleration reaches the teams who need it depends less on the tooling and more on whether those teams have the organizational authority to act on what the tools surface.

Same mistake, different decade

The industry keeps calibrating risk models to the threat landscape that existed when the model was built. EPSS, CVSS base scores, patch SLAs — all of these reflect a world that’s already a few years in the past. AI-assisted exploitation is changing the denominator: the cost and skill required to move from disclosure to weaponization. When that denominator drops, everything built around the old denominator needs recalibration.

None of this means the old advice is wrong. Know your inventory. Prioritize by exposure. Patch fast. Those principles hold. What’s changed is the acceptable timeline for executing on them.