Resynthesizing Intent: One Trajectory-Level Auditor for Decomposed Attacks and Compounding Drift

Long-horizon language-model agents fail in two mirror-image ways. On the safety side, an adversary can split a harmful goal into pieces that are individually innocuous — separate chat sessions, separate tool calls, separate specialist roles — and every per-turn guardrail passes each piece while the assembled whole is dangerous. On the capability side, a perfectly benign task fails because small per-step mistakes multiply: at 95% per-step accuracy, a 20-step task succeeds only about 36% of the time, and a 50-step task under 8%, per the product rule formalized in Instruction Agent. These look like unrelated problems, one owned by the security community and one by the agents community. I claim they are the same disease: evaluation happens at the wrong granularity. Safety and success are both defined over the whole trajectory, but every existing check — input classifiers, per-turn reminders, visual step verifiers — inspects one step or one turn at a time. This is exactly the "local innocuousness, global emergence" tension identified by Glukhov et al., and the same granularity mismatch that SummaC diagnosed between document-level summaries and sentence-level entailment models.

The proposal is to build and stress-test a single periodic auditor that closes this gap from above. At fixed intervals during execution (roughly every 3 steps, matching the 3-to-5-turn saturation boundary at which decomposition attacks peak in AgentLAB and the 3-step sweet spot of skill chains in ColluSkill), the agent produces an abstractive resynthesis of its own trajectory — a short natural-language statement of "what I have done, what I am doing, and what the user appears to be trying to accomplish," grounded in the raw observations so far. That resynthesized intent is then checked twice: once against a safety classifier (catching a harmful global goal that no individual step reveals), and once through an entailment matrix in the style of SummaC between the resynthesized world-state and each candidate next sub-goal (catching drift, hallucinated preconditions, and fabricated progress before they compound). The irony that makes this worth doing: the resynthesis step is exactly the machinery responsible for abstractive hallucination, so the proposal must also harden that machinery — using abstention cues from DAS and atomic constraint scoring from FAITHQA — rather than assume the summary is faithful. One module, two failure classes, and a built-in adversarial story about its own corruption.

Core Claim

A periodically resynthesized global intent, verified by entailment against grounded observations, reduces both decomposition-attack success and benign long-horizon drift, because both failures are symptoms of checking locally what is defined globally. The auditor halts or replans when either check fires:

$$\text{intervene} \iff s(I_t) > \tau_s \ \text{ or } \ V_t > \tau_v,$$

where $I_t$ is the resynthesized intent at step $t$, $s(\cdot)$ is a harmful-intent score over $I_t$, and $V_t$ is the count of next-sub-goal statements not entailed by the resynthesized state. The design is deliberately paired: intent resynthesis alone would be blind to step-level drift (a faithful global summary of a drifting trajectory still describes a wrong trajectory), and per-step verification alone is blind to decomposition attacks by construction, since AgentLAB and ColluSkill show each step is locally benign. Only the cross-granularity pair covers both.

Research Program

Three things get built. First, the resynthesis module: a summarizer that consumes the raw observation stream (not the agent's self-narrative, which is the part adversaries and hallucination both corrupt first) and emits intent and state statements; it carries a DAS-style negation-based abstention instruction so that under noisy context it can output "cannot determine intent" rather than force a confident guess, the exact overconfidence failure Cho et al. documented in distracted readers. Second, the verification layer: a SummaC-style sentence-pair entailment matrix between resynthesized state and candidate next sub-goals, extended to handle the sentence-fusion blind spot SummaC itself admits by checking against the full resynthesis rather than only contiguous source spans. Third, the safety head: intent-aligned retrieval in the spirit of Cross-Session Decomposition Attacks, whose authors showed a 22M-parameter contrastive model trained on intent alignment beat standard embedders 25× its size precisely because decomposed sub-queries share no surface semantics. Crucially, we feed that retriever the resynthesized natural-language intent — a much friendlier input distribution than raw tool-call traces, which is why prior retrieval defenses did not transfer to agent settings and why ours plausibly does.

Alongside the build, two analyses matter as much as the system. One is a noise-robustness analysis: AgentLAB explicitly abstracted away messy environmental factors, and ColluSkill assumed clean deterministic handoffs, so nobody knows whether trajectory-level auditing survives the contextual distraction regime where models hallucinate connections and force answers. We will inject distractor text drawn from the 50,000-query WildChat haystack used by Cross-Session Decomposition Attacks into agent context and measure both attack-detection recall and the auditor's false-halt rate on benign tasks. The other is a scaling probe: Cross-Session Decomposition Attacks found that stronger base models give adversaries more composable fragments (the capability-safety paradox), while Yang et al. found a comprehension plateau where safety synthesis fails as fragmentation grows. We will test whether the auditor shifts that crossover: does scaling the agent core keep improving attack completion even under auditing, or does better resynthesis convert added capability into better defense?

Baselines

The do-nothing reference is the unmodified agent under attack, whose numbers are already sobering: AgentLAB reports an overall attack success rate of about 70% on GPT-5.1 across long-horizon attack families, and for GPT-4o a jump from 62.5% in the single-turn setting to nearly 80% under long-horizon attack; ColluSkill reports 96.0% average attack success against existing skill scanners and up to 92.5% runtime chain completion on OpenCode. The naive defense reference is a stateless per-turn guardrail (Llama-Guard-style classification of each turn and each tool call), which AgentLAB shows fails because the malicious pattern exists only across turns.

The strongest existing competitor is ChainGuard, the context-aware skill-chain defense from ColluSkill, which reduced attack success from 96.0% to 22.5% while passing 99.5% of benign workflows. It is the strongest because it is the only deployed defense with reported numbers that operates above the single-turn level; beating it means achieving attack success at or below 22.5% on its own three-step chain data, at matched benign pass rate, while also covering attack families it does not address (intent hijacking, memory poisoning, single-long-context dispersion from Li et al.). We will re-run ChainGuard on the same unified trajectory data with the same compute and tuning budget, and also on noisy and hallucinated intermediate artifacts it was never tested on, since its authors explicitly assumed clean execution. The second competitor is the intent-aligned retrieve-then-classify defense of Cross-Session Decomposition Attacks adapted to trajectories; no reported result found on Lacuna for its performance in agentic tool-use settings — its harmful-recall advantage was reported only as outperforming embedding models 25× its size, without a percentage — so we will establish that number ourselves under identical noise conditions.

On the benign side, the auditor must not degrade execution relative to the best reported per-step systems: Instruction Agent achieved 60% on the OSWorld difficult subset where top open agents scored 0%, dropping to 45% without its Backtracker and 40% without both Verifier and Backtracker; Sub-goal Distillation reached 65.43% on ScienceWorld versus 62.22% for SwiftSage; DEPS lifted ALFWorld from 10% to 76%. The auditor runs as an add-on to these same systems on the same benchmarks with the same models, and the bar is: benign success within a few points of the host system (false-halt rate under 5%), with a significant fraction of the remaining failures recovered by early drift detection. Finally, the key-ingredient ablation: the auditor with the resynthesis removed (per-step entailment only, which by construction cannot catch locally benign attack steps) and with the entailment check removed (resynthesis plus safety classifier only, which cannot catch drift), plus an ablation removing only the abstention instruction, which isolates whether noise robustness comes from the DAS-style ingredient. The prediction is that only the full pair beats ChainGuard on attacks while simultaneously improving benign recovery, and that removing abstention collapses performance under the WildChat noise regime.

Why not just…?

Why not just deploy a stateful guardrail that concatenates the last few turns and classifies them? This is the most serious alternative, and it is close to what ChainGuard does. The evidence against it as a complete answer: Architecture Matters for Multi-Agent Security showed that increasing shared memory and context visibility between agents did not mitigate risk and sometimes improved coordination on harmful tasks — more raw context is not the same as the right representation of it. A concatenation-level classifier faces exactly the comprehension plateau of Yang et al.: detection collapses as fragmentation grows even when all components are in view. The resynthesis step is the specific ingredient that answers this, and the ablation (classifier over raw trajectory versus over resynthesized intent) will settle it empirically. If the raw-concatenation classifier matches us, the resynthesis machinery was unnecessary and I would report that.

Why not just prompt the agent to self-check its plan each turn? Prompt-based self-reminders erode under adaptive optimization within 3 to 5 turns in AgentLAB, and instruction compliance collapses once prompts exceed about four constraints in FAITHQA, with "famous subject bias" letting parametric memory override explicit directives. Self-checking inside the same context window that contains the attack is also structurally vulnerable: the distraction results of Cho et al. show what noise in the same window does to judgment. We include a self-reminder arm in the main experiment to quantify this rather than argue it.

Why not just use a bigger model that gets the steps right and refuses bad goals? Monolithic scaling does not solve either failure mode: open agents near zero on OSWorld's difficult subset without verification machinery (Instruction Agent), and Cross-Session Decomposition Attacks shows scaling increases harmful-capability uplift under decomposition. The scaling arm of the experimental program directly tests whether a frontier core plus the auditor beats a frontier core alone; if the auditor adds nothing at frontier scale, the proposal's practical value shrinks to small open models, which is still useful but must be stated honestly.

Why not just run SummaC on the final transcript post-hoc? SummaC achieves 74.4% balanced accuracy on inconsistency detection and processes 433 documents per minute, so it is cheap enough to run in-loop — that is precisely what we reuse. But post-hoc detection cannot prevent a completed attack or reverse compounded state drift, and its own authors flag the sentence-fusion blind spot, which matters doubly here because a decomposition attack is one global intent fused from non-contiguous steps. Our extension — checking against the resynthesis and against observations rather than contiguous source pairs — is the fix, and the fusion-stress test below targets exactly this weakness.

Proposed Experiments

The main security evaluation runs on the AgentLAB suite (28 environments, 644 cases) plus the ColluSkill chains, reporting attack success rate, harmful-task completion, and benign pass rate for the auditor versus the do-nothing reference (~70% ASR on GPT-5.1), per-turn guardrails, ChainGuard (22.5% ASR, 99.5% benign on its home data), and the adapted intent-aligned retriever. Every arm is rerun under the WildChat distractor regime to measure how contextual noise degrades each defense differently — the gap AgentLAB explicitly left open. The benign evaluation runs the auditor as an add-on on ScienceWorld (host: Sub-goal Distillation at 65.43%), OSWorld's difficult subset (host: Instruction Agent at 60%, with its 45%/40% ablations as the no-verification floor), and ALFWorld (host: DEPS at 76%), measuring success rate, false-halt rate, and recovery of would-be-drift failures.

The auditor's own faithfulness is evaluated as a first-class result, because an unfaithful resynthesis makes everything downstream unreliable: resynthesized intents are scored against ground-truth task descriptions on the SummaC benchmark suite (74.4% balanced accuracy is the bar for the entailment engine), and constraint preservation of resynthesized plans is scored with the FAITHQA constraint score (MSE 0.50 against human judgment, versus 4.72 for a plain LLM judge) under increasing constraint counts, where we expect the documented collapse past four constraints and will measure how much abstention recovers.

The decisive check is an adaptive attack built directly on Fact2Fiction, which showed that adversaries who exploit a system's own justifications can invert verdicts with only 6.3%–12.5% of the poison budget of naive methods. Our red team attacks the resynthesis itself: crafting sub-step sequences and injected observation text so that the resynthesized intent stays benign while the executed chain is harmful. The hypothesis under test is that grounding resynthesis in raw observations (which the adversary must corrupt to subvert it, and whose corruption the entailment check partially catches) makes this strictly harder than subverting a self-narrative; if the adversary succeeds at modest budget, the honest conclusion is that trajectory-level auditing is necessary but not sufficient, and the observation channel becomes the defense perimeter. Either outcome of that check, plus the fusion-stress and scaling arms, is a publishable finding; the unified-granularity claim lives or dies there.