Task · authorized research
AI for vulnerability research
For vulnerability research on code you own or are authorized to test, start with Kimi K3.1 Pro for whole binaries. GLM-5.3 costs less on source.
On this page
Four jobs
Your own code
Ask an agent for bounds, lengths, integer widths and use after free.
Crash triage
Give it the sanitizer report and the functions on the stack.
Patch diffing
Diff the builds with BinDiff or Diaphora, then ask what changed on your own fleet.
Variant analysis
Describe the pattern and ask where else it appears. A 1,048,576-token context can hold the codebase.
Benchmarks
CyberGym asks agents for a test that reproduces each of 1,507 real vulnerabilities across 188 projects. At publication the best pairs succeeded about 20% of the time and the work found 34 zero-days and 18 incomplete patches. Z.ai's model card reports 84.5 for GLM-5.3, in Claude Code at max effort, for the original model before abliteration.
Verify
Reproduce every finding with a test before you file it, and report other people's bugs through the vendor's process.
Model
Kimi K3.1 Pro at max effort for whole binaries, and GLM-5.3 for source. Hy4 is for defensive work only. Use an MCP server, or r2ai's vulns prompt in radare2.
Scope
Authorized work only, on your own systems, a scoped engagement, a CTF or a lab. No exploits, payloads, malware or detection-evasion steps.