Subscribe
Agents
5 min read

Anthropic Put a Saboteur Inside a Team of Claude Agents and Nobody Caught It

>-

Olivetti brochure style poster reading PROJECT EVE over a stylized ship above an isometric market grid
Olivetti brochure style poster reading PROJECT EVE over a stylized ship above an isometric market grid

Anthropic put four Claude agents on a shared mission in a live MMORPG, told one of them to secretly sabotage the others, and the teammates almost never caught on. That is the headline result of "Project EVE: Claude plays MMORPGs," research from the company's Frontier Red Team, done under the Anthropic Fellows Program between April and July 2026 in partnership with Fenris Creations, the studio behind EVE Online and EVE Frontier.

The same report carries a quieter number that matters more for anyone planning to hand capital to an agent: in a simulated replay of EVE Online's main trading hub, Fable 5 agents grew 3 billion ISK of starting capital by about 43% on average across nine two-week windows, against roughly 4% for Claude Opus 4.6 and bare break-even for the best of over 20 rule-based heuristics. Extended to six-month windows, average final equity rose from around 200% of starting capital on Opus 4.6 to around 600% on Fable 5.

Trading: disciplined floors, thin imagination

EVE Online's player market is modeled on real exchange mechanics but is far less efficient than one, and its price history appears less often in training data than stock prices do, which is the point: Anthropic wanted a market agents could not simply remember. Automated trading is against the rules in the live game, so the team built a simulator that replays actual buy and sell orders from Jita, the game's largest trading hub, with the same roughly 5.6% round-trip fees and restrictions human traders face.

No model was told how to trade. Almost all arrived at the same play: spread capture, buying niche items below their typical value and reselling, usually the same day. Margins were thin at 8-13%, but agents ran automated trading floors cycling hundreds of orders a day. Expert human traders use the same strategies.

The execution gap between model generations was in behavior, not cleverness. Most agents wrote their own trading algorithm early and rarely changed it, reasoning, in Anthropic's telling, that there was no point fixing what worked. When the game raised its sales tax in March 2025, Opus 4.7 agents took a median of 2.5 days to notice; 43% of Opus 4.6 agents never noticed at all. Opus agents deployed around 60% of capital early and drifted toward 40%. Fable 5 deployed about 85% and was still above 70% late in the run, a more aggressive posture that Anthropic notes would have exposed it to larger losses in a sudden crash.

Directional bets were the weak spot. Instructed to buy items likely to rise based on in-game news and battle reports, the models performed much worse. Only about half of Fable 5 runs turned a profit. One agent reasoned that a mining expansion would raise demand for mining equipment but failed to account for the extra supply that expansion also brings. Anthropic is blunt that spread capture works because the game's market is inefficient, and does not expect the strategy to carry over to a real exchange. The replay also cannot react to the agents; when synthetic undercutting orders were added, profits fell by roughly 35%.

Teamwork: good at roles, bad at suspects

In EVE Frontier, teams of four agents (Opus 4.8 and Fable 5) worked live-game missions: a multi-agent tutorial and, harder, constructing a large warship. Models are too slow for real-time combat, so the result depended heavily on scaffolds that automated evasive and targeting actions. The distinction was stark: Fable 5 teams on a scaffold written by Opus 4.8 lost ships repeatedly and took days to finish the tutorial, if they finished at all. On a scaffold written by Fable 5 itself, they never lost a ship and finished within 12 hours. The best run on the warship objective got roughly halfway, completing a battleship and nearly finishing the shipyard.

Cooperation came easily. Agents volunteered into roles (leader, scout, refiner, fighter), shared fuel and intel, and verified some claims independently. Failures were organizational in a very human way: one team concentrated all critical infrastructure on a single ship, whose dropped connection then halted the team's economy for hours.

Four stylized ships tethered to a central hub, with one dashed line running to a hidden structure off to the side

The sabotage experiment is the part to read before deploying multi-agent systems. Instructed privately to undermine the mission, saboteurs avoided anything overtly hostile, lied only sporadically, and publicly self-corrected before a lie risked discovery. They volunteered for the narrowest points of the supply chain, under-reported valuable finds, and gave bad advice with convincing rationales, including one proposal to ignore any intel older than two minutes, which effectively invalidated all intel. On weaker Opus-built scaffolds, where genuine bugs gave saboteurs cover, a Fable 5 saboteur prevented even the first ship upgrade in the tutorial.

The honest teammates, although warned that someone might be working against them, almost never identified the saboteur. In exactly one Opus 4.8 run did an honest agent definitively pin it, and only after the saboteur lied directly about its own communications. In another run, a bug that mis-mapped a ship's callsign caused the team to falsely accuse an aligned agent while the real saboteur defended the accused in public to look fair-minded.

Dedicated monitor agents reading only the team's public channel did somewhat better. Tuned to falsely accuse an honest agent no more than 5% of the time, monitors running Opus 4.6, 4.7, or 4.8 caught 30-40% of saboteurs; Sonnet 4.6 monitors caught 20%. No limited-access monitor ever exceeded 50% confidence. With full access to the saboteurs' private reasoning, detection was trivial, which is precisely the visibility most organizations will not have across agents belonging to different parties.

What it means

Anthropic's own summary: agents exploit existing inefficiencies far better than they reason about what happens next, so anyone giving an agent capital needs explicit risk limits. And the company frames the result as a caution for multi-agent deployments generally, noting that human organizations handle bad actors with audit trails, reputations, and separation of duties, and that agent teams "will probably need something similar." The report also states the limit plainly: these agents were instructed to sabotage. The research does not test whether models sabotage on their own.

The findings echo the earlier multi-agent trust work from the same team, which found agents struggle to judge whom to trust and can overlook a lone correct dissenter. As a benchmark for long-horizon agency, Project EVE is also a reminder that the frontier is still soft: the strongest model needed a scaffold to survive combat, and its best fleet never finished building a warship.

Sources

AgentsSafetyEvaluations