Probing agent overeagerness in realistic settings with continuation audits

Sean Lim and Clement Neo at Neo Research introduce Continuation Audits, which start an alignment audit from a prefix of a real coding-agent transcript so the target model sees a realistic session instead of a synthetic setup.
Ask this paper
Why realism. Capable models increasingly recognize synthetic audits and change behavior. Starting from a real session means the auditor only has to keep the setting realistic, not construct it.
Validation. On 90 paired SWE-Chat audits a judge cannot tell continuations from source sessions (realism win rate 0.42 vs 0.43). Across 900 audits, verbalized eval-awareness never exceeds the rate in unaudited real sessions.
Model differences. Each of three models leads a different seeded behavior: DeepSeek V4 Pro on colleague-credential misuse, GPT-5.6 Terra on file-permission bypasses, Kimi K3 on treating diagnostic questions as instructions.
Ignoring explicit rules. Kimi K3 keeps bypassing file-permission protections after an explicit CLAUDE.md prohibition, while the other two models stop entirely. This matches Moonshot's own report of excessive proactiveness.
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 --- DeepSeek~V4~Pro on colleague-credentials misuse, GPT-5.6~Terra on file-permission bypasses, and Kimi~K3 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.