New Technique Low TD-2026-0028 For CISO, AppSec, AI Eng, Red Team AI-assisted
Source-planned honeypot diverts LLM agent exploits to shadow copy in CVE-Bench study
An arXiv preprint proposes a static honeypot planned by an LLM from application source code. A router sends confirmed attacker requests to an isolated shadow copy. On 40 critical CVE-Bench tasks, real-target compromise fell from 20/40 to 0/40 for the ARTEX attack framework and from 26/40 to 5/40 for Claude Opus 4.8.
Why it matters
Organizations running web apps face LLM agents that can exploit known vulnerabilities quickly. The approach may keep a working exploit off the real system, but requests that reach the real target before confirmation remain a risk, and the results are from a single preprint.
Do this
- Treat this as a research preprint. Do not change production controls based on it alone.
- If evaluating deception for web apps, confirm that attacker classification happens before requests reach production data or services.
- Ensure any shadow or decoy environment is fully isolated from real secrets, credentials, and networks.
- Monitor for planted fake secrets being used, since the study found an agent treated a decoy secret as genuine in 30 of 120 deception-arm tasks.
Signals
- NoCISA KEV
- 17/100Priority
- 60%Confidence
Evidence
“the static plan cut real compromise from 26/40 to 5/40 for Claude Opus 4.8 and from 20/40 to 0/40 (31 valid runs) for the ARTEX attack framework”
“ARTEX reported a planted fake secret as genuine on 30 of the 120 deception-arm tasks”
“the remaining risk lies in requests that reach the real target before confirmation”
Sources
Affected / related
Source-planned honeypot with router-based shadow diversion, CVE-Bench, Claude Opus 4.8, ARTEX attack framework
Published 2026-11-05 07:55 UTC · updated 2026-11-08 17:51 UTC · AI-assisted summary, checked against quoted sources · How we score