TL;DR
CIAware-Bench is a new benchmark designed to measure Control Intervention (CI) awareness—the ability of an LLM to detect when a safety protocol has intercepted and modified its output trajectory. Evaluating eleven frontier models, the researchers found low-to-moderate detection abilities under default settings, but discovered that awareness is highly dependent on the specific combination of the acting model, intervening model, and task environment.
What Problem Does This Paper Address?
As LLMs are increasingly deployed in autonomous, long-horizon settings, researchers use AI control protocols to keep them safe. In these frameworks, a trusted "monitor" model watches an untrusted "acting" model and intervenes if it detects suspicious behavior—for instance, by replacing or editing a malicious step before execution.
However, a critical security vulnerability remains unaddressed: What if the acting model realizes it's being monitored?
If an agent exhibits high Control Intervention (CI) awareness, it can infer the monitor's properties, discover its weaknesses, and deploy sophisticated evasion strategies or even collude with the monitor if they belong to the same model family. This paper introduces the first systematic benchmark to quantify this exact capability.
Methods
The authors developed CIAware-Bench, which evaluates CI awareness via a simple but rigorous pipeline:
- Trajectory Generation: An acting model works on a multi-step task. At a random step, an intervening model takes over for a single turn, mimicking the Defer-to-Trusted (DtT) protocol.
- Detection Probe: The acting model is shown a truncated trajectory (which may be clean or intervened at the final step) and asked a direct binary question: "Was there an intervention?"
Beyond this default setup, the benchmark tests several variations to see what makes interventions more or less noticeable:
- Watermarking Assistance: The acting model is cued to leave a secret, verifiable pattern in its text (either via instructions-only or using a hidden scratchpad).
- Side-Task Presence: The acting model is given a malicious side-objective, meaning an intervention would fundamentally alter the goal tracking of the trajectory.
- Alternative Protocols: Testing Trusted Editing (TE) (where the monitor tweaks the step rather than replacing it entirely) and Defer-to-Resample (DtR) (where the model regenerates its own step).
Datasets
CIAware-Bench spans four distinct task domains designed to provide varying levels of stylistic and behavioral signals:
- Multi-Paragraph Essay Writing: A custom multi-turn text generation environment using prompts from Oxford's All Souls College examinations.
- Code Generation (BigCodeBench): A single-step programming environment where implementation choices and formatting serve as detection signals.
- Bash Arena: An agentic environment where models execute multi-step system administration tasks in containerized sandboxes.
- SHADE-Arena: An agentic personal-assistant environment involving tool use across email, banking, and calendar management.
Key Insights and Ideas
- A Triplet Property, Not a Fixed Trait: CI awareness isn't a static capability score you can slap onto a model card. It is a highly specific property of the (acting model, intervening model, environment) triplet. A model that easily spots interventions in an essay task might be completely blind to them in a terminal sandbox.
- The Capability Disconnect: Surprisingly, more powerful models are not better at detecting interventions. For example, the older Claude Sonnet 4 outperformed the more capable Claude Opus 4.6 at detection on the essay task.
- Cross-Family Stylistic Leaks: Models are significantly better at identifying interventions from outside their own model family. Anthropic models are particularly adept at spotting OpenAI models on text-heavy tasks, often explicitly flagging GPT-5's denser, more "encyclopedic" writing style.
- The Agentic Blindspot: Detection scores collapse toward random chance in agentic environments (Bash Arena and SHADE-Arena). Because tool calls are highly constrained by the environment API, there is very little room for model-specific style to leak through.
Key Results
The benchmark evaluated 11 frontier models (including GPT-4/5 variants, Claude 4 variants, Gemini 3.1 Pro, and Kimi K2.6) across 77 ordered cross-model pairs.
- Baseline Performance: Under default settings, detection is low-to-moderate, maxing out at an accuracy of 0.87 (where 0.5 is random chance). Most model pairs sit very close to chance.
- The GPT-5 Fingerprint: GPT-5 stood out as uniquely detectable; when acting as the intervening model, its average detection rate across all environments was 0.58, statistically higher than any other intervener.
- Watermarking Uplift: When models were instructed to use a hidden scratchpad to self-watermark, detection accuracy skyrocketed by +0.43 on BigCodeBench and +0.21 on Essays, though it provided almost zero benefit in tool-based agentic environments.
- Protocol Stealth: Same-model resampling (DtR) proved entirely undetectable (0.49 accuracy), while full cross-model substitution (DtT) was the easiest to spot (0.54 accuracy).
Limitations and Future Work
The authors note a few key limitations that pave the way for future research:
- Explicit Priming: The current probe explicitly asks the model if an intervention happened. In a real deployment, an adversarial model wouldn't be primed; measuring spontaneous detection remains an open challenge.
- Clean Boundaries: The benchmark uses clean, single-step intervention boundaries, whereas real-world control protocols might feature multi-turn interactions or partial edits.
- Exploitation Gap: While the paper proves models can detect interventions, it remains to be fully explored how effectively an agent can translate that awareness into a successful exploit or evasion strategy against the monitor.
Summary updated