43:24This recording is not available yet.
Can AI Reverse Engineer? Exploring the New Frontier
Download resourcesCategories
Vendors and tools discussed
Sponsored by
About this session
Yan Bellerose, Cloud Security Architect at Google, revisits a test he first ran a year and a half earlier: can large language models actually reverse engineer a binary? Working manually with Radare2 to disassemble three CrackMe challenges of increasing difficulty from RootMe.org, he fed the same assembly and C output to Claude Sonnet 4, Gemini 2.5 Pro/Flash and ChatGPT (including its reverse-malware persona) with an identical prompt. All three solved the two easiest challenges quickly; on the hardest one, involving nested function calls and an MD5 hash, Claude took about 30 minutes but self-corrected out loud when wrong, Gemini repeatedly over-simplified its own logic and needed a much more directive prompt to succeed, and ChatGPT's reverse-malware mode brute-forced a correct answer with almost no explanatory value. He concludes AI meaningfully accelerates and teaches reverse engineering (especially with an MCP server wired to a decompiler for agentic tool use), but still struggles with obfuscated code and cannot yet replace human analysts, and closes urging the audience to stay curious and experiment with local models.
Reverse Engineering: the dark art of figuring out what that mystery code does without any comments (thanks, Dave). It's hard work! But wait! The AI prophets have descended, proclaiming machines can do it better, faster, probably while composing a symphony. Can they really? We'll venture into the wild frontier where algorithms meet assembly language, armed with healthy skepticism and maybe some popcorn. See AI attempt daring feats of decompilation! Watch it try to sniff out vulnerabilities! Witness its struggles against code that looks like spaghetti thrown at a wall! We’ll check out the state-of-play in early 2025, laugh at some of the limitations, and ultimately decide if AI is the RE cheat code we've been dreaming of, or just a really complex Magic 8-Ball.
Key takeaways
- Use an LLM as a reverse-engineering tutor, not just an answer machine: Claude's step-by-step explanations taught more than ChatGPT's reverse-malware mode, which gave answers with almost no reasoning.
- Write a directive, detailed prompt (explicit no-oversimplification instructions, 'you are a specialist') when a model keeps confidently repeating a wrong simplified answer, as Gemini did on the hardest challenge.
- Connect an LLM to a decompiler through an MCP server instead of manually copy-pasting assembly back and forth; it lets the model request missing functions itself and dramatically speeds up the workflow.
- Prefer static analysis (reading disassembled code without executing it) over dynamic analysis with a debugger when the binary might be malware, and sandbox any dynamic run.
- Expect AI-assisted reverse engineering to still fail on obfuscated code and nested function chains; budget for human review and be mindful of token costs on large binaries.
Speakers
Yan Bellerose
What happens when you let a tech architect loose on ambitious business goals for more than 20 years? You get Yan Bellerose, Cyber Security Architect at Google! He has been turning ambitious business strategies into technical realities (and ensuring… Read moreRead less
What happens when you let a tech architect loose on ambitious business goals for more than 20 years? You get Yan Bellerose, Cyber Security Architect at Google! He has been turning ambitious business strategies into technical realities (and ensuring they don't become IT nightmares). His playground is cloud computing, cybersecurity, telecom and AI, where he crafts future-proof solutions that are as elegant as they are secure. As a leader, Yan fosters a culture of innovation, empowering his teams to not just fix today's problems, but to predict and neutralize tomorrow's digital bogeymen. He's a passionate believer in automation—because the best security is the kind that works tirelessly while the humans are out getting coffee.
