New Technique Low TD-2026-0080 For SOC / IR, Red Team, CISO AI-assisted
Shibboleth traps fingerprint LLM version and harness behind agentic cyberattacks
The paper introduces shibboleth traps, modular honeypot services that elicit discriminative behaviors from agentic attackers without requiring a successful prompt injection. Combined declarative and behavioral inference distinguishes among 15 candidate models with an average 90% top-1 accuracy. In a separate experiment across four harnesses, fixed probe rules identify the correct harness with 92% accuracy.
Why it matters
Defenders collecting telemetry from attacker interactions could infer which LLM and agent harness an attacker uses, supporting attribution and tailored response. The paper also notes that agents instructed to conceal identity still leave behavioral evidence.
Do this
- Log full request sequences and timing from honeypot and decoy services so behavioral patterns can be analyzed later.
- Evaluate deploying decoy services that return structured responses to agentic clients as a research or telemetry source.
- Review how your tools mask secrets in tool output and enforce command time limits, since the paper shows harness behavior can be inferred from these effects.
- Treat fingerprinting results as probabilistic and corroborate them with other indicators before using them for attribution.
Signals
- NoCISA KEV
- 17/100Priority
- 60%Confidence
Evidence
“The traps make informative interactions a condition for obtaining resources the agent seeks, eliciting model declarations and characteristic writing without requiring a successful prompt injection.”
“combined declarative and behavioral inference distinguishes among 15 candidate models with an average 90% top-1 accuracy”
“fixed probe rules identify the correct harness with 92% accuracy”
Sources
Affected / related
shibboleth traps, harness fingerprinting
Published 2026-11-09 13:27 UTC · updated 2026-11-12 18:05 UTC · AI-assisted summary, checked against quoted sources · How we score