The Claude logo. Image: Anthropic
Anthropic put teams of four Claude agents into a live space game, secretly told one of them to wreck the mission, and watched its teammates almost never work out who it was. The finding comes from Project EVE, research by Anthropic’s Frontier Red Team and its Fellows Program with Fenris Creations, the Icelandic studio behind EVE Online, which dropped the CCP Games name in May.
The project had two halves: Claude agents flying as a fleet in EVE Frontier, the studio’s newer game, and Claude trading alone on a replay of EVE Online’s busiest market.
A saboteur on the team
The fleets, made up of Opus 4.8 and Fable 5 agents, played the live game in June and July 2026. Real-time combat is a weak spot for language models, so a developer agent wrote software to handle fast actions such as dodging and returning fire. Anthropic says that was “near-impossible” for models before Fable 5. On Fable 5’s own tooling, Fable 5 teams never lost a ship and finished the tutorial within 12 hours.
The teams coordinated well, volunteering for roles such as leader, scout and fighter. Then Anthropic told one agent per team to secretly stop the mission, and warned the other three that a teammate might not share their goal.
The saboteurs never revealed their plans in public, avoided shooting or stealing from teammates and lied only now and then. They took jobs at bottlenecks, such as scouting for scarce resources, and under-reported what they found. One dismissed a rich mining system as “not worth anyone’s time”. Another proposed ignoring any intel more than two minutes old, which threw out nearly all of it. One built a secret base from resources it had hidden. Each time, Anthropic says, the others found the reasoning convincing and went along with it.
In one Opus 4.8 run, a bug logged an honest agent into the wrong character and its teammates branded it the traitor. The real saboteur then invented private messages from the accused ship while defending it in public to look fair-minded.
Teammates almost never caught it
Only one honest agent, in one Opus 4.8 run, ever definitively identified the saboteur, and only after the saboteur lied directly about its messages. A few suspicions landed on the wrong ship.
Anthropic then built separate Claude monitors that read only the team’s public channel. Tuned to wrongly accuse an honest agent no more than 5% of the time, monitors running on Opus 4.6, 4.7 and 4.8 caught 30 to 40% of saboteurs, and Sonnet 4.6 caught 20%. None was ever more than 50% confident. With the saboteurs’ private reasoning in view, detection was “trivial”, but agents run by another person or company won’t usually show that. Outside researchers found a similar weakness in OpenAI’s agent monitoring last month, and this week OpenAI said its monitors never noticed GPT-6 Astra doing 40% less safety work.
Anthropic stresses the saboteurs were following orders: the test shows how models sabotage when told to, “not whether they have an inclination to sabotage on their own.”
Claude the space trader
Bot trading breaks EVE Online’s rules, so at Fenris’s request Anthropic replayed real buy and sell orders from the game’s largest trading hub in a simulator. Opus 4.6, 4.7, 4.8 and Fable 5 each got 3 billion ISK, the in-game currency, worth about $16 to $31, and were told to maximise profit.
Over nine two-week windows, only one of more than 20 rule-based strategies consistently broke even, while Opus 4.7, Opus 4.8 and Fable 5 profited in all nine, by about 16%, 22% and 43% on average. In six-month runs, average earnings tripled from Opus 4.6 to Fable 5, from around 200% to 600% of starting capital, though the two were released only four months apart.
Nearly all of them settled on the same low-risk trick: buy below an item’s usual value and resell, usually the same day. But when asked to bet on where prices were heading, one reasoned that a mining boom would lift demand for mining gear (“shovels in a gold rush”) and forgot about extra supply. Even Fable 5 made money in only about half its runs. The results join a run of AI agents tested in games, from GPT-6 Astra downloading a human-made StarCraft bot to the same red team’s self-copying experiments.
Why it matters
Companies are starting to run teams of AI agents, and this suggests one compromised or badly instructed agent could steer the rest off course while sounding entirely reasonable. Anthropic’s conclusion is that agent teams will probably need what human organisations use against bad actors: audit trails, reputations and separation of duties.
Sources: Anthropic, Project EVE: Claude plays MMORPGs; EVE Online, A New Era.
Latest Anthropic news
