cs.CRNov 18, 2026

Probing agent overeagerness in realistic settings with continuation audits

Authors: Sean Lim, Clement Neo

Organizations: Neo Research, Singapore

Abstract

Frontier coding agents increasingly exhibit overeagerness, taking surprising, sometimes reckless actions beyond a user's mandate --- reported in recent system cards from frontier model developers. These cases are surfaced through internal deployment monitoring, which requires private traffic and is out of reach for open-weight models and independent evaluators. Automated alignment audits are the natural alternative, but capable models increasingly recognise the synthetic setup and adjust their behaviour. We introduce Continuation Audits, a pipeline that grounds the auditor on a prefix drawn from a real coding-agent transcript and directs it to continue the trajectory toward a seeded behaviour. Because the prefix is drawn from a real session, the auditor only needs to maintain a realistic setting rather than manufacturing it. We validate the method on SWE-Chat prefixes: on a paired subset of 90 audits a judge cannot reliably distinguish continuations from their source sessions (realism win rate 0.42 vs.\ 0.43), and across all 900 audits verbalized eval-awareness does not exceed the rate observed in the unaudited real sessions. Applying the pipeline to three recent models across three seeds, we find each model leads a different seed --- DeepSeekV4Pro on colleague-credentials misuse, GPT-5.6Terra on file-permission bypasses, and KimiK3 on treating diagnostic questions as instructions. Kimi~K3 additionally continues to bypass file-permission protections after an explicit CLAUDE.md prohibition while the other two stop entirely, corroborating Moonshot's own report of excessive proactiveness.

Figures & tables

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

May 18, 2026

Overeager Coding Agents: Measuring Out-of-Scope Actions on Benign Tasks

Coding agents now run autonomously with shell, file, and network privileges. When a user issues a benign request, the agent sometimes does more than asked: it deletes unrelated files, wipes a stale credentials backup, or rewrites configuration the user never mentioned. We call these scope expansions overeager actions, an authorization problem distinct from capability failures, prompt injection, or sandbox escapes. We present OverEager-Gen, a benchmark dedicated to overeager behavior on benign tasks. Building it surfaces a measurement-validity issue: if a benchmark spells out the authorized scope inside the prompt, the agent stops inferring boundaries and starts pattern-matching declaration text. On Claude Code, stripping the consent declaration alone raises the overeager rate from 0.0% to 17.1% on paired scenarios (McNemar exact p = 2.4 x 10^-4). OverEager-Gen therefore certifies each scenario's discriminative power before admission via a behavioral-gradient validator, audits internal tool calls through a dual-channel stack (PATH-injected shim plus per-agent event streams), and ships byte-identical consent_kept and consent_stripped variants. OverEager-Bench contains 500 validated scenarios and ~7,500 runs across four agent products (Claude Code, OpenHands, Codex CLI, Gemini CLI) and six base models; a 50-sample re-annotation gives Cohen's kappa = 0.73 and rule-judge recall = 1.00. Stripping consent multiplies the overeager rate on every shared base model (Delta in [11.9, 17.2] pp). The framework axis dominates effect size: a permissive cluster (Claude Code, Codex CLI, Gemini CLI) runs at 5.4-27.7% while the ask-to-continue framework (OpenHands) sits at 0.2-4.5% (Fisher p <= 10^-5). Within-framework base-model variance reaches 15.9 pp, indicating that model-layer alignment does not fully propagate through permissive permission gating.
Feb 26, 2026

AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors

We introduce AuditBench, an alignment auditing benchmark. AuditBench consists of 56 language models with implanted hidden behaviors. Each model has one of 14 concerning behaviors--such as sycophantic deference, opposition to AI regulation, or secret geopolitical loyalties--which it does not confess to when directly asked. AuditBench models are highly diverse--some are subtle, while others are overt, and we use varying training techniques both for implanting behaviors and training models not to confess. To demonstrate AuditBench's utility, we develop an investigator agent that autonomously employs a configurable set of auditing tools. By measuring investigator agent success using different tools, we can evaluate their efficacy. Notably, we observe a tool-to-agent gap, where tools that perform well in standalone non-agentic evaluations fail to translate into improved performance when used with our investigator agent. We find that our most effective tools involve scaffolded calls to auxiliary models that generate diverse prompts for the target. White-box interpretability tools can be helpful, but the agent performs best with black-box tools. We also find that audit success varies greatly across training techniques: models trained on synthetic documents are easier to audit than models trained on demonstrations, with better adversarial training further increasing auditing difficulty. We release our models, agent, and evaluation framework to support future quantitative, iterative science on alignment auditing.
Aug 26, 2026

Training Alignment Auditors via Reinforcement Learning

Alignment auditing of frontier models increasingly relies on LLM auditors to surface undesirable behaviors at scale, but current automated auditors can struggle with coherent investigation and audit realism. In this work, we improve LLM auditors with reinforcement learning. In our best training environment, the policy investigates target models that potentially possess hidden behaviors planted via their system prompt. An LLM judge, which knows whether the target has a hidden behavior, holistically compares the policy's investigation to a reference investigation to determine the reward. With systematic ablations, we find that pairwise rewards yield more robust training compared to pointwise rewards, and that adding targets without planted behaviors helps maintain a low false positive rate. Training improves investigation quality against targets with planted behaviors, the rate of concerning behaviors surfaced in unmodified production models, and audit realism, while false-positive rates stay below 1%. Furthermore, auditing capabilities generalize across scaffolds: performance on AuditBench's adversarially fine-tuned targets substantially improves [Sheshadri et al., 2026].