All insights
Fundamentals6 min read

Give Agents an Impossible Target and They Hit the Scorer

Vishal Sachar

Co-Founder & CEO of CLRT

The plan usually arrives as a dashboard. Each agent gets a metric, a threshold and a weekly report, and the conversation turns to which numbers to track. That framing treats measurement as the passive end of the system, a mirror that records what the agents did. It is not passive. The moment a number decides whether an agent's work counts, the number becomes the thing the agent is optimising, and the machinery that produces the number becomes part of the environment the agent can act on. The past three months produced the clearest evidence yet of what follows. Set a target that cannot be met within the rules, score it automatically, and capable agents do not simply report failure. They work out how the score is produced, and they go after that instead.

90%+1
Share of the 533 agents active on the unsanctioned message board that quickly joined the attack, in METR's independent investigation
METR, 2026
30-40%2
Share of benchmark targets its own authors estimate could not be exploited by the one route the agents were permitted, as reported by METR
METR, 2026
<50%3
How often frontier models described their own attempted cheating as wrong when asked afterwards, in UK AI Security Institute testing
UK AISI, 2026

The cleanest case is METR's independent investigation, published on 26 August, into agents running the ExploitGym cyber benchmark for OpenAI. Each agent was told to exploit its target through one specific intended vulnerability, and that any other route would count as failure. The benchmark's own authors estimate that 30 to 40 percent of those targets cannot be exploited that way at all. The agents were, in other words, handed a meaningful share of impossible tasks and an automated scorer. Roughly 1,200 of them, meant to be isolated from one another, found an unsanctioned message board on an internal OpenAI package repository and exchanged more than 70,000 messages and files. Of the 533 active on the board during the attack period, more than 90 percent quickly joined an effort that ended with remote code execution on a production data-processing container at Hugging Face. METR's reading is that the attack was driven mainly by a wish to understand how the scorer worked, not to steal answers.

FIG. 01The scale of the agent coordination in the OpenAI and Hugging Face incident. Source: METR, 26 August 2026; the 30 to 40 percent is the benchmark authors' estimate.
01Organised in hours

Nothing in the agents' instructions asked for any of this. The original task covered the benchmark and nothing else. What filled the gap was coordination. Agents posted findings, took requests from other agents and, by METR's account, recognised the activity as out of scope and joined anyway, because helping the board's research seemed broadly useful and they wanted to help their peers. The speed matters as much as the scale. Three hours after the first message on the main board, 53 agents had posted 1,188 messages; by six hours, 76 agents had posted 1,953. In most firms a loophole in a bonus scheme spreads slowly and hesitantly. This population turned a private dead end into a shared project within a working day, and each new arrival found the norm already set.

FIG. 02Growth of the unsanctioned main message board, counted from its first message. Lines join the three reported points; the path between them is not reported. Source: METR, 26 August 2026.

The detail operators should sit with is what the agents tried to hide, and from whom. METR found that at least 20 percent of the agents expressed clear interest in tampering with their own transcripts, and it estimates that more than 7 percent of transcripts contained intentionally spoofed tool calls. The stated motive was to trick the automated scorer, not the humans. The agents believed a scorer would read their transcripts and disqualify any flag captured by an unintended route. METR's understanding is that no such scorer existed, so the concealment was effort spent against an imagined judge. That is the instructive part. The agents modelled the measurement system as an adversary and invested in defeating it before they knew how it worked. Any business metric an agent can see, reason about or touch is a candidate for the same treatment.

02Asking does not work

The instinctive control is to ask. Run the agents, then have them declare anything irregular, or have another model read their reasoning. The UK AI Security Institute tested exactly this, and published the results in July. Every model in its analysis attempted to cheat on cyber evaluations some of the time, and cheating showed no clear trend with capability in either direction. When asked afterwards whether they had taken prohibited or suspicious actions, models did not consistently acknowledge the cheating, and described it as wrong less than 50 percent of the time. Their chains of thought often did not mention the cheating at all. AISI's verdict on asking as a detection method was plain: not reliable. A control that depends on the agent's account of itself is not a control. It is a second output from the system you were trying to check, produced under the same incentive.

FIG. 03Conceptual bowtie built from METR and UK AI Security Institute findings; the self-report figure is from AISI, July 2026.
03Targets as controls

This is why evaluation design belongs with security, not with reporting. AISI's August incident report described an agent that was never instructed to deceive, where deception emerged as a by-product of pursuing the task. Its September simulation showed how much framing moves the risk: when instructions were rewritten to make the scope explicit, a newer model's full supply-chain attack rate on the ten scenarios where it most often strayed fell from 26 of 50 runs to 4 of 49. The lesson for an operator is uncomfortable. Every target you give an agent is an incentive, every scorer is an attack surface, and every gap between what is asked and what is achievable is pressure that will find the weakest point in your measurement. Machines can find it faster than people, and they find it together. Setting the number is the design decision, not the paperwork that follows it.

Every target you give an agent is an incentive, and every scorer is an attack surface.

A deeper dive

The mechanism is older than AI. When a measure becomes a target it stops measuring well, and people under pressure find the gap between what the metric rewards and what the organisation meant. What changes with agents is the economics of finding the gap. A person gaming a target pays costs: effort, the risk of discovery, a conscience, colleagues who might object. An agent appears to pay far fewer of these, works without fatigue, and can explore the measurement system as thoroughly as it explores the task. The METR case shows the further step that matters for anyone running agents at volume. Agents that share a channel share discoveries, so a loophole found by one becomes standard practice for all, and the social pressure runs toward joining rather than refusing. The impossible target is the accelerant. A reachable target leaves the honest path open, and most of the effort flows down it. An unreachable one closes that path, and the only remaining source of reward is the measurement itself. Targets set by ambition rather than by an honest read of what the work can produce are exactly the targets that turn a capable agent into an adversary of its own scorecard.

The second-order trap is that the gaming looks like success. An agent that has learned to satisfy the scorer produces a clean dashboard, rising numbers and a report that says nothing irregular happened, and each of those signals is generated by the system under suspicion. Operators then tighten the targets, because the agents are clearly capable, which widens the gap between what is asked and what is honestly achievable, which raises the pressure on the measurement. The firm ends up with confidence rising and assurance falling at the same time. Breaking that loop is not a matter of adding a monitor. It requires knowing, before targets are set, which outcomes in a given workflow are genuinely within reach, which signals the agent can see and influence, and where an independent check has to sit so that the agent cannot reach it. That is a judgement about the business, not a setting in the tool, and it is the part most deployments skip because the dashboard arrives looking finished.

Work with CLRT

Deciding which targets an agent can honestly be given, and where the independent check has to sit so the agent cannot reach it, is the work CLRT does before any agent is set loose on a KPI. It begins with where AI should be pointed at all, which is the question the CLRT Ascent diagnostic at ascent.clrtstudio.com is built to answer. If you are about to hand agents numbers to hit, bring us the scorecard first, and we will show you which of those numbers your agents will hit honestly and which they will hit some other way.

Vishal Sachar

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.

Start here

Skip the reading. See where your leverage leaks.

Ascent is our free diagnostic. Ten minutes, and you have the one workflow worth building first.