AI attacks move in minutes, so yearly pentests are no longer enough

The pressure point. Snyk's data puts the problem in numbers: 218 percent more lines of code per developer, and 62 percent of LLM output insecure or broken. Even if vendor-defined, the direction matches what leaders already see: agent-generated code lands faster than security review can absorb it. The result is a backlog growing faster than remediation. This tension frames the talk.

The accelerated attacker. The talk claims AI attackers chain low-severity issues into critical exploits. One cited model breached 600 firewalls across 555 countries; that number should be treated with suspicion. The useful signal is speed: average successful AI attack in 24 to 34 minutes, fastest in four, and 43 percent of MCP servers carrying vulnerabilities. None of this is independently verified, but the direction is plausible: automated attackers do not wait for the next pentest.

The broken baseline. Cohen walks through the classic stack: SAST is cheap but blind to runtime issues like broken object level authorization; DAST throws payloads at endpoints but misses business logic; human pentesters understand context but cost too much, run once or twice a year, and deliver PDFs developers must triage. That model was designed for slow release cycles. With agents shipping code continuously, an annual test leaves hundreds of days uncovered.

The product pitch. Snyk EVO is presented as an LLM-native replacement: an orchestrator plans, a recon agent maps endpoints and APIs, specialized agents test vulnerability classes, a judge validates exploitability, and remediation plus reporting close the loop. Differentiators are continuous execution on every pull request and context from SAST, DAST, and other scanners fed into the LLM. That context-first claim is the most credible part. It is also a vendor pitch.

The buying criteria. The most reusable content is four questions for any AI pentest vendor: is it continuous, does it reason about application context, what context can it consume, and does it prove a bug is exploitable or just flag it. That is a useful procurement checklist. The tradeoff is unstated: continuous agent pentesting on every code change will generate signal, but also cost, noise, and remediation pressure. The architecture is sensible; the evidence is not measured.

What to watch. Engineering leaders should treat this as an early signal, not an imperative. The gap between point-in-time pentests and continuous attacks is real, but Snyk presents no independent benchmark, production failure data, or false positive rate. Pilot a bounded service, track whether reported exploits are real and fixable, and compare against current review costs before committing. The problem is urgent; the specific product remains unproven.