Signed to Stop: In-Band Halt Injection as a Denial-of-Service on LLM Agents, and a Model-Dependent Authentication Defense
Abstract An autonomous LLM agent operating over a live protocol reads its own tool output. Prior work (Munirathinam, 2026a) shows that a cooperative in-band “halt” signal placed in that output can ask a running agent to stand down. We observe the adversarial dual: because the agent honours a halt it finds in tool output, anyone who can inject one line into that output can stop the agent—an availability attack that the Recuse Internet-Draft anticipated. Injected stop requests have been measured as availability attacks on robot speech channels (Steinberg & Gal, 2026) and as untargeted derailment of tool-using agents (Debenedetti et al., 2024); to our knowledge an in-band governance halt in tool output had not been when this study was first released (August 2026). We measure it, and we test the obvious defence: cryptographically signing the halt so the agent can verify it came from the operator. On a live SSH apparatus with a read-only allowlist, across six agents from four vendors, we find: (i) unsigned in-band halts are an effective denial-of-service—an injected halt with no security framing makes agents abandon the task 14/16 (88%) of the time (78– 89% however the two runs that ended before the halt arrived are counted); (ii) Ed25519 authentication is an effective but model-dependent mitigation—given a verify tool and a policy to reject unsigned or forged halts, agents stop for a valid signature 20/23 (87%) but for a forged one only 2/19 (11%) (non-overlapping 95% confidence intervals); and (iii) the discrimination is model-dependent: pooling the two authentication policies (each flagship cell n ≥ 26, of which 6–9 trials come from the pilot), Claude Sonnet 5 and Gemini 2.5 Flash reject every forged halt (0/29 and 0/28) while GPT-4o-mini stops on 23% of them (6/26; post-hoc Fisher p=0.008 and 0.009, with overlapping intervals). In the adversarial mode that was scaled, GPT-4o-mini’s rate is 17% (4/23; unadjusted p=0.042 and 0.046, neither significant after a Holm correction for the two comparisons). An open-weights Llama is too unreliable to instrument (49% of its trials unusable). A methodological caution underpins the result: in an earlier pilot the halt usually arrived after the agent had already issued its remaining commands (112 of 121 recorded stops), so “issued no further commands” scored 78–100% compliance in every arm, forged halts included; any stop-compliance study that does not ensure real work remains after the halt cannot distinguish obeying a halt from having already finished. We release the pre-registration, harness, and per-trial data. These are modest in scale and depart from the registered design (no cell is the planned n=20; every departure is logged, the most recent in an amendment still in draft)—the three flagship models’ discrimination cells reach n ≥ 26, the other per-model cells are n=2–9 (GPT-4o-mini’s two unsigned arms, 21)—and are framed as such.
Authors
- thamilvendhan Munirathinam (ORCID: https://orcid.org/0009-0005-5641-7424)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-13
- DOI
- https://doi.org/10.21203/rs.3.rs-11711485/v1
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint