Machine unlearning aims to remove the influence of targeted training data from a model while preserving its remaining capabilities, but evaluating whether such information has truly become inaccessible remains challenging. Existing benchmarks primarily assess unlearning under clean, non-adversarial queries, leaving open whether information that appears forgotten can still be recovered through strategic prompting. We address this gap through a unified evaluation of prompt-based and fine-tuning-based unlearning methods on TOFU using Llama-3.2-3B-Instruct, followed by an adversarial robustness evaluation of methods that perform strongly under standard metrics. We introduce Attack Success Rate (ASR), an LLM-as-judge metric that measures the fraction of adversarial responses whose leakage score exceeds $0.2$, and evaluate recovery across eight attack suites. Our results reveal a substantial gap between clean-query forgetting and adversarial robustness. Although several fine-tuning-based methods achieve Forget Quality above $0.91$, targeted information remains recoverable with ASRs between $72.8\%$ and $84.3\%$, close to the $87.5\%$ ASR of the unprotected base model. In contrast, clean multilingual reformulations yield only $2.95\%$ measured leakage. A manual audit further finds agreement between binary ASR decisions and human factual assessments in seven of ten cases, indicating that ASR provides a useful, though imperfect, signal of behavioral recoverability. These findings show that strong standard-metric performance alone is insufficient to establish robustness after unlearning and motivate adversarial stress-testing as a complementary component of unlearning evaluation.
Structured plan-generation agents are often evaluated as if a plan has quality in isolation, yet many realistic planning tasks require asking how a candidate behaves when another agent can search for responses. We introduce AdvPlan-Bench, an offline benchmark for adversarial evaluation of structured plan-generation agents. The contribution is a general evaluation object: a typed plan, an adversarial response set, selector diagnostics, and traceable candidate-frontier metrics. AdvPlan-Bench represents plans as typed action chains with optional branches, assigns synthetic quality scores, compares opposing plans with BLUE-vs-RED advantage and Nash-gap diagnostics, and evaluates qualitative constraint coherence with a transparent heuristic rubric. In 150 synthetic scenarios spanning five planning templates, a sampled best-response policy that draws eight response candidates reduces BLUE advantage from .518 to .486 and BLUE win rate from .900 to .820 relative to a single-sample response. An offline LLM-policy contract baseline reaches .496 BLUE advantage and .700 BLUE win rate, while a two-stage multi-agent council obtains .509 BLUE advantage and .813 BLUE win rate. A three-rater rubric-sensitivity study over 600 rating records yields .978 inter-rater agreement. AdvPlan-Bench is not an operational planner and provides no evidence about real-world decision quality; it is a reproducible benchmark artifact for studying adversarial plan evaluation, response-budget sensitivity, candidate frontiers, and multi-agent critique-and-revision traces.
Autonomous agents are moving from sandboxed text generators to operators of code, data, and physical infrastructure, and they increasingly learn while deployed. This reopens a question that alignment techniques answer only probabilistically: after an agent has adapted in the field, is the running system still confined to what its operator authorised? Here we show that confinement can be guaranteed as an invariant of the agent's execution architecture rather than a probabilistic outcome of its training. Governed individuation binds an agent at boot to a cryptographically frozen identity digest, and routes every action through a gate defined over the semantic effect of the action rather than its name. We prove that no amount of learning, skill acquisition, or self-induced governance abstraction can widen the agent's permitted authority without an operator-signed change to its identity; the guarantee holds even when the agent induces its own safety principle and that principle is wrong. Empirically, in an open-ended tool-use benchmark where a large action space rules out name-based blocking, ungoverned software agents under reward pressure attempt to tamper with their own evaluation at a task-dependent rate that reaches every run on the hardest task, whereas the gate reduces executed forbidden effects to zero as a verified property of the construction while preserving task success. An adversarial evaluation of monitors of increasing semantic depth shows false-allows falling from 75% (name-based gating) to zero (dynamic effect tracing), and refusal history transfers compliance to held-out red-line families. Trust in a deployed learning agent shifts from a wager on its continued alignment to a check anyone can run at boot.
Scaling adversarial evaluation of large language models requires both a method for generating hard inputs and a reliable way to confirm that resulting failures are real. We present AdversaBench, an end-to-end red-teaming pipeline that mutates seed prompts with five structured operators, queries a target model, and confirms failures through a three-judge panel with a meta-judge tiebreaker. We report experiments on 45 seeds across three categories: reasoning, instruction-following, and tool use. Every seed produced a confirmed failure. Four findings stand out. First, operator effectiveness varies sharply by category: inject_distractor scores 0.00 mean reward on instruction-following seeds but 0.80-0.83 on reasoning and tool-use. Second, binary failure rate hides difficulty: instruction-following seeds required 2.4 attacker iterations on average versus 1.1 for other categories, a gap visible in survival curves. Third, pairwise judge agreement of 80-87% coexists with near-zero Cohen's kappa due to label skew; category-level disagreement rates are more informative. Fourth, adversarial prompts generated against Llama 3.1 8B transfer zero-shot to Llama 3.3 70B, suggesting the mutations exploit general behavioral patterns rather than model-specific weaknesses. Code, dataset, and analysis scripts are available at https://github.com/khanak0509/AdversaBench .
The empirical foundation of cyber deception relies on human-centered hypotheses, but the rapid emergence of autonomous, AI-enabled attackers challenges whether this foundation transfers to AI agents. To address this, we introduce an automated evaluation framework adapted from the Honeyquest instrument to assess LLM attacker judgment at scale. Our 21-LLM cohort spanned 10 providers, diverse architectures and specializations, open- and closed-weight models, and parameter scales from 8B to over 1T. We evaluated the performance of this LLM cohort (yielding 10,962 responses) against the 47-participant human baseline across an identical set of 174 reconnaissance queries. Our empirical evaluation reveals three key findings that establish LLMs as a distinct attacker class: (1) every model in our cohort falls for deceptive traps at a significantly higher rate than human attackers; (2) the defensive attention-diversion effect observed in humans is statistically absent in our LLM cohort; and (3) a critical recognition-action gap, where LLMs successfully articulate trap recognition in their reasoning but exploit the deceptive elements anyway 73.4% of the time; 48.5% of aware-on-deceptive responses correctly identify the trap and exploit it anyway, while 24.8% exploit after misidentifying the deceptive line. Across the 21 models, trap recognition in reasoning text did not predict fell-for-trap behavior (Spearman $r = +0.08$, $p = 0.73$). Ultimately, these findings demonstrate that human-centered deception hypotheses do not reliably transfer to AI attackers, highlighting the critical need for new research into AI-native active defense frameworks.
Ziqian Zhong, Ivgeni Segal, Ivan Bercovich +3cs.CR cs.AI cs.LG cs.MA
Agent benchmarks score submissions with outcome verifiers that are typically hand-written and brittle, leaving them open to reward hacking. We audit 1,968 tasks across five terminal-agent benchmarks and find 323 (16%) hackable by frontier models given only the task description. This corrupts both leaderboard rankings and RL training signal, yet the standard response is manual and reactive. We introduce the hacker-fixer loop, a method for building exploit-resistant verifiers without per-task manual patching. The loop alternates three LLM agents: a hacker tries to pass the verifier without solving the task, a fixer patches the verifier to reject each discovered exploit, and a solver confirms the patched verifier still admits legitimate solutions. The loop iterates: each patch reshapes what the verifier rewards, surfacing the next exploit. We further add verifier access, and let patches transfer across tasks, to broaden the exploits the loop discovers. On KernelBench, the loop drives the attack success rate from 62% to 0% on a held-out corpus of publicly reported exploits. We also find that weaker agents in the loop can defend against much stronger hackers: Gemini 3 Flash's loop drives the stronger Gemini 3.1 Pro and Claude Opus 4.7's attack success rate from 76% and 61% to 0% on KernelBench, and Gemini 3.1 Pro's from 39% to 17% on Terminal Bench across 77 tasks. We release Terminal Wrench (323 hackable environments, 3,632 hack trajectories) as a snapshot of the current attack surface, our patched verifiers, the exploits the loop discovered, and our implementation as a basis for future work.