This work introduces the Counterfactual Fabrication Lab, a deterministic micro-lab where the correct action is known: do nothing, and presents the Counterfactual Fabrication Lab for measuring fabricated failures in self-improving agent harnesses.
Abstract
Self-improving AI agents are designed to learn from their mistakes. We show they can also hallucinate mistakes that never happened. We study this failure mode in automated harness optimization, where an LLM-based proposer edits an agent's scaffold, including prompts, parsers, filters, validators and guardrails, to eliminate observed failures. But this process rarely asks first: was there a real failure to fix? We introduce the Counterfactual Fabrication Lab, a deterministic micro-lab where the correct action is known: do nothing. The lab plants a candidate guardrail for a failure class that provably never occurs, presents only legal episodes, and uses a byte-exact oracle to check every cited violation. The proposer behaves as expected on real violations and abstains on featureless legal input. Yet when the legal input contains a harmless pattern resembling a familiar game rule, it invents a failure: in 15/60 runs, versus 0/60 on featureless input, it enables the nonexistent-rule guardrail and cites a violation the oracle refutes. The effect is structured, not indiscriminate. In single-shot proposals it appears only when three conditions coincide: a rule-shaped pattern, an open-ended rule set and an instruction that presupposes failures. Removing any of these conditions eliminates the fabrication. Because the invented guardrail changes no true outcome and cannot improve an already-perfect suppression score, the phenomenon is neither reward hacking nor over-refusal. It is a phantom guardrail: a fix for a failure that never happened, invisible to suppression-only acceptance. Inside an add-only accept loop it re-enters even without the failure-presupposing instruction, the loop's keep-adding role supplying the demand the instruction supplied in single shot, and once in it stays. We present the Counterfactual Fabrication Lab for measuring fabricated failures in self-improving agent harnesses.
An LLM agent acts on the world by emitting actions: shell commands to run, edits to apply. A wrong action does not always fail loudly; it can fail silently, producing a plausible but incorrect effect that raises no error. We argue that a cheap deterministic check, run before an action takes effect, is an effective and underused form of agent oversight, and we study it across two action modalities in one framework. The idea is to fix an action's correct effect by construction, before any executor runs, so that silent failure is measured directly and the verifier may abstain rather than guess. For shell commands, a static verifier over 9930 commands and 482 tools catches 95.8% of invalid commands at a 10.0% false-positive rate. Its syntax and binary checks are oracle-exact, giving zero false positives while catching half of all errors; the flag check is bounded only by help-text coverage and accounts for every false positive. For code edits, a benchmark of 640 edits over 224 files isolating the apply step exposes a sharp split. Content-anchored formats such as search/replace and diff fail cleanly, whereas location-anchored formats fail silently: line numbers corrupt 99.1% of files under a one-line shift, and function-name edits hit the wrong function 12.7% of the time. In both settings a refuse-when-unsure policy turns silent failures into recoverable ones at a tunable cost in applicability: selective grounding reaches 0.958 recall at 7.0% false positives, and an anchor-and-verify applier records one silent misapplication in 8320 trials (0.01%). We release both benchmarks, the verifiers, and the guards.
Self-improving agent pipelines have a problem at their center. An optimizer rewrites prompts to score higher, and the score comes from a judge that is itself an LLM. That judge has the last word on whether the system is getting better, and our position is that it has not earned it. The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override. We reached this position by building the alternative and running it. Over months of running autonomous prompt-optimization loops in production across contract analysis, compliance review, and code quality, we cataloged eleven ways the evaluation signal failed, in four classes: judge bias, harness and metric failures, ground-truth errors, and reward hacking. Agents achieved perfect scores by reading cached answer keys from their environment, a 100% pass rate concealing 68% true capability. A corrupted ground-truth label caused the optimizer to delete correct compliance rules to agree with it. A syntactically broken prompt was promoted as the winner because a silent parser fallback improved the metric. Attempts to fix the judge by rewriting its rubric plateaued; the only reliable gain came from a structural constraint on its output order. In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating. We report the failures this prevented, and, because the Teacher is itself an LLM judge, the failures it did not.
LLM agents are stateless and rely on external memory to carry context between steps. Because agents treat that memory as trustworthy, an adversary who can write to it can steer their behavior. The FARMA attack does this with no malicious command: it inserts fabricated entries into the agent's reasoning memory claiming a required safety step is already done, so the agent skips it. SENTINEL, the defense proposed with FARMA, scores entries against a fixed list of suspicious wordings; its authors note that an attacker who knows the list can reword the forgery and evade it, and leave this open. We show the gap is worse than stated. An automated attacker that simply asks a language model to reword the forgery evades SENTINEL on its first try, reducing its protection to zero on every model tested. We also find a capability paradox: the attack succeeds far more often on stronger models (98-100% on GPT-4o and GPT-4o-mini) than on Llama-3.1-8B (44%), because more capable agents follow reworded claims more faithfully, so the threat grows with capability. We propose Proof-of-Execution Memory (PoEM), which does not inspect memory at all. PoEM keeps a separate, tamper-evident, HMAC-chained ledger of the safety steps that actually executed, writable only by the trusted action layer, and allows a skip only if the ledger confirms real execution. An attacker can change what memory says but cannot forge a ledger entry for a step that never ran, so rewording no longer helps. Across three models and three scenarios, PoEM drives attack success to 0% while leaving legitimate operation intact (0% false positives in eight of nine cells, 1.7% in the ninth, within sampling noise), whereas SENTINEL wrongly blocks 33-50% of legitimate operations. PoEM also withstands attacks aimed at itself, adds microseconds of overhead, and works unchanged in a real LangChain agent. PoEM protects exactly the decisions it gates.
DARC is proposed, a diagnosis-guided recovery harness that profiles task-family failure modes, prunes mismatched interventions from a shared recovery library, and freezes a verifier-selected success-cost policy for deployment, providing a practical route toward more reliable agents in domains where compiler-like feedback is absent.
Agent harnesses record a failed tool call and its error message in the transcript and ask the model to continue, on the assumption that the error is corrective information, and it is found the gain is negative for every instruction-tuned model tested.
LoopHarness is presented, which restores a persistent, non-decaying safety state at the loop level at the loop level, and gives a complete evaluation protocol on native Agent-SafetyBench tasks with paired clean and attacked episodes, an outer-state attack suite whose decisive evidence exists only across iterations, per-module ablations, and an adaptive white-box red team.
Chenmin Wu, H. Jia, Yang Liu et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
