DeepDive - From APEG (2008) to AI Agents (2026) - Old-School Exploitation, New-Scale Automation
AI-generated exploits are making headlines, but the underlying technique is nearly two decades old. Back in 2008, researchers from Carnegie Mellon University and UC Berkeley demonstrated that security patches could be automatically transformed into working exploits. Their paper, Automatic Patch-Based Exploit Generation is Possible: Techniques and Implications, introduced APEG, a system combining program analysis, symbolic execution and constraint solving to generate exploits from differences between vulnerable and patched software. The researchers successfully generated exploits for five Microsoft programs using Windows Update patches, sometimes within minutes.
In October 2026 Epoch AI demonstrated that the open-weight GLM 5.3 model could autonomously exploit the patch gap across six recent Firefox releases, at a marginal cost of approximately $1,000 per targeted release.
The concept remains remarkably similar:
2008 — Traditional APEG
Vulnerable binary + security patch → program analysis → exploit generation
2026 — AI-powered exploitation
Public security fixes + AI agent → vulnerability analysis → iterative exploit development → working exploit
So, What Actually Changed?
The original APEG approach relied on specialized program-analysis techniques and targeted particular vulnerability classes. Modern AI agents can combine source-code analysis, tool execution, debugging and iterative experimentation across more complex attack scenarios. Epoch's Firefox experiments went beyond reproducing a single patched vulnerability: the agent combined multiple vulnerabilities, including one previously undisclosed flaw, to bypass browser security boundaries. However, the result is not universal. Epoch's researchers did not achieve equivalent success against Chrome, and the reported $1,000 cost excludes substantial initial research and harness-development expenses.
Why Does It Matter?
The fundamental security problem has not changed: a published security fix can reveal a vulnerability before the fix reaches every user. What changes with AI is the potential to lower the expertise barrier and automate more of the exploit-development process.
Summary: AI didn't invent patch-based exploit generation. It may make an old, well-understood attack technique accessible to many more actors. The real challenge remains the same as in 2008: reducing the dangerous window between security-fix disclosure and deployment.