AI-Assisted Reverse Engineering: What the Evidence Shows
Published tool documentation, reporting on decomp.dev, and DARPA competition results show specific ways AI assists software analysis. Each source below is linked to the claim it supports.
1. An Open-Source Bridge to Binary Analysis
The GhidraMCP project documentation describes an MCP server for Ghidra that lets compatible clients request decompilation, list symbols, and rename methods and data. These are documented tool capabilities, not evidence that a specific agent reconstructed an entire product or produced an exploit.
2. What Changed at decomp.dev
On October 15, 2026, decomp.dev's operator announced that the public project index and new submissions would stop while existing project pages and historical data would remain available. The operator cited AI-assisted forks and rushed ports as reasons, and 404 Media covered the decision. The announcement describes a change to the community index; it makes no claim that scraping overwhelmed the infrastructure.
3. What the Media Reports Actually Cover
- Functional reimplementation: Ars Technica reported on AI-assisted alternatives intended to imitate Adobe applications, while noting that the projects were unfinished. Its reporting concerns replicated behavior and interface, not a demonstrated extraction of Adobe source code.
- Vulnerability research: WIRED covered AI systems finding software vulnerabilities. The article concerns security testing, not a verified source-code cloning incident.
- Espionage activity: The Record reported Anthropic's account of a Russia-linked group using Claude during a cyber-espionage campaign. It does not report that this group reconstructed software from a released binary.
- Contested source-code allegations: The New York Post reported Runlayer's lawsuit alleging that Rippling misused confidential material, including source code shared during a partnership. These are allegations, and its follow-up reported Rippling's denial and counterclaims. The case is distinct from reverse engineering a public release.
4. DARPA AIxCC: A Measured Result
In the 2025 AI Cyber Challenge final, competitors' autonomous systems analyzed more than 54 million lines of code, found 54 of 63 synthetic vulnerabilities, and patched 43. These figures come from DARPA's results announcement. The competition measured vulnerability discovery and patching in open-source codebases; it did not measure recovery of proprietary source code from released binaries.