New Technique Medium TD-2026-0076 For AppSec, AI Eng, CISO AI-assisted
ActiveSecHarness: neurosymbolic reference monitor for agentic coding harness security
The paper presents ActiveSecHarness, a reference-monitor architecture for AI coding harnesses that couple language models to shells, filesystems, and network sockets. It combines HMAC capability tokens across four trust rings, a 2D product taint lattice with a provenance ledger, and deterministic Ring-0 OS and tool guards. On 252 fixtures, Ring 0 reports ABR of 1.000 and FPR of 0.000 at 12.7 microseconds per check.
Why it matters
Agentic coding harnesses can turn injected instructions in retrieved content into misuse of dual-use binaries such as bash, curl, and find. Results are from a single arXiv preprint on a custom fixture set, so they may not transfer to production environments.
Do this
- Inventory agent harnesses that grant shell, filesystem, or network access and list which binaries they can invoke.
- Restrict agent tool permissions to least privilege and deny shell and network access by default for retrieved-content workflows.
- Enforce deterministic allowlist or policy checks at the tool-execution layer rather than relying on regex blacklists or model-based judges alone.
- Log agent tool invocations and secret-bearing file writes so provenance can be reviewed during incident response.
- Validate any reference-monitor approach against your own benign and attack workloads before relying on reported error rates.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“turning the harness into an automated Living-off-the-Land (LotL) platform: once an adversary injects instructions into retrieved content, dual-use binaries such as bash, curl, and find become the payload”
“Static regex blacklists miss structural evasions, whereas LLM-as-a-Judge guardrails incur high latency while over-blocking benign work”
Sources
Affected / related
Living-off-the-Land via AI coding harness, ActiveSecHarness, qwen3:4b
Published 2026-11-10 11:52 UTC · updated 2026-11-12 17:35 UTC · AI-assisted summary, checked against quoted sources · How we score