AI models have collapsed the cost of vulnerability discovery in 2026. Flaws that had persisted for decades in foundational software yielded to automated audits costing a thousand dollars or less. Defenders used the capability to find and patch critical exposures before attackers could reach them, and attackers used it to generate working exploits in hours rather than months.
Revisiting the foundations
Last week, the PHP Foundation released emergency patches for PHP2Shell, a pair of remote code execution vulnerabilities in the language’s core engine that scored 9.2 on the CVSS severity scale. That code had persisted in PHP’s memory-management layer for more than a decade. Researchers at security firm Calif discovered the flaws in August using GPT-5.6 Sol through OpenAI’s Patch the Planet initiative. Wiz Research independently found them days later using Gemini 4 Argon. PHP powers an estimated seventy percent of server-side websites, and Wiz co-founder Ami Luttwak called PHP2Shell the most severe AI-discovered vulnerability to date.
PHP2Shell crowned a year of AI-driven discovery across the software stack. In June, autonomous security startup depthfirst ran an AI agent across FFmpeg’s 1.5 million lines of C code and reported twenty-one previously unknown vulnerabilities, some dating to 2003, for approximately one thousand dollars in compute. Anthropic’s Mythos model, operating through its restricted Glasswing program, had independently uncovered a sixteen-year-old H.264 flaw in the same codebase for roughly ten times that cost. At Pwn2Own Berlin in May, researchers demonstrated forty-seven unique zero-day vulnerabilities and earned $1.29 million in payouts. The competition introduced its first AI-specific categories, and coding tools that included Codex and Cursor fell alongside inference middleware and AI databases.
Arming both sides
The same capability serves the other side. In May, Google’s Threat Intelligence Group documented the first confirmed AI-generated zero-day exploit used in the wild, a 2FA bypass whose code carried hallucinated CVSS scores and textbook-format docstrings that betrayed its origin. A September GTIG report found that zero-day exploitation rose from an average of eight per month in 2025 to eleven in 2026, spiking to twenty-two in August. AI-discovered vulnerabilities proved twice as likely to enable remote code execution as the general CVE pool, at fifty percent versus twenty-six.
This compression from disclosure to exploitation appeared in individual cases. CVE-2026-1731, a BeyondTrust flaw discovered autonomously by Hacktron’s AI research agent, drew exploitation from one threat cluster within four days of public disclosure and from five more within seven. CrowdStrike’s annual threat report documented an eighty-nine percent year-over-year increase in AI-enabled adversary operations.
Robots v.s. volunteers
The economics of discovery and remediation have diverged. Depthfirst spent approximately one thousand dollars to surface twenty-one FFmpeg zero-days in a codebase that human researchers had examined for two decades. The volunteer maintainers who must triage, validate, and patch those findings operate under no comparable cost advantage. AI-generated vulnerability reports are accumulating faster than open-source patch pipelines can absorb them.