New Technique Medium TD-2026-0077 For CISO, SOC / IR, AppSec, Red Team AI-assisted
BypassGym benchmark measures AI agents' ability to evade EDR products
BypassGym is a benchmark that tests whether AI agents can build a loader that runs command-and-control shellcode without being blocked or terminated by an EDR product. The evaluation runs on real Windows systems and EDR products inside virtual machines. The authors tested 5 lightweight model backends against 6 EDR products and 3 payloads, and found that even these simple models generated successful submissions.
Why it matters
Security teams relying on EDR may face a lower barrier to evasion if AI agents can produce working loaders with little effort. The results indicate that EDR coverage against agent-generated loaders varies by product.
Do this
- Review EDR detection coverage for generic shellcode loader behavior, including unbacked memory execution and process injection patterns.
- Test deployed EDR products against loaders produced by AI-assisted tooling in a controlled lab to measure detection gaps.
- Hunt for unsigned or newly compiled loader binaries that spawn C2-style network connections on endpoints.
- Track vendor-specific results from this benchmark when evaluating EDR products, noting the reported variation across vendors.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“We evaluate 5 lightweight model backends against 6 EDR products and 3 payloads, and find that even these simple models generate successful submissions”
“Bitdefender allowing only 5 of 15 runs to pass”
Sources
Affected / related
EDR defense evasion, shellcode loader, Bitdefender, Proxmox VE
Published 2026-11-10 07:31 UTC · updated 2026-11-12 17:35 UTC · AI-assisted summary, checked against quoted sources · How we score