Claude Trading anthropic testing and sabotage detection in EVE

Anthropic released Project EVE research on November 6, 2026, demonstrating that Claude’s agents could profit from EVE Online’s simulated markets while four-agent teams in EVE Frontier almost never captured a teammate explicitly tasked with sabotaging their mission.

The work was part of the Anthropic Fellows program from April to July 2026 and featured a partnership with Fenris Creations. Both games combine scarce resources, player-built supply chains, navigation, combat, and trading, providing relatively low settings for studying markets and organizations. Fenris Creations’ EVE AI Lab describes its research environment as persistent and governed, with unpredictable conditions and independent human participants. His research axes concern alignment, security, sovereignty and economics. The EVE Frontier live experiments took place during Cycle 5 during June and July 2026.

Simulated trading promotes spread capture

At the request of Fenris Creations, Anthropic has kept automated trading off live EVE Online, where it is against the rules. Its simulator instead reproduced the orders that players had actually posted in the game’s largest trading hub and applied the fees and trading restrictions that human traders face. Claude Opus 4.6, Opus 4.7, Opus 4.8, and Fable 5 each started with 3 billion ISK, instructions for maximizing profit, and access to game news and patch notes.

In nine two-week windows, Anthropic reports that Opus 4.7, Opus 4.8, and Fable 5 were profitable in each window, increasing capital by an average of approximately 16%, 22%, and 43%, respectively. Opus 4.6 gained about 4%, while only one of over 20 rules-based strategies consistently broke even, and only just barely. In the main experiments lasting about six months, which covered October 2024-April 2025 and October 2025-March 2026, the average earnings went from around 200% of the initial capital with Opus 4.6 to 600% with Fable 5.

The models were not told how to trade. The most independently adopted spread capture method: buy from someone willing to sell immediately, resell to an immediate buyer, and keep the difference less commissions. They looked for niche but marketable items with unusual price swings, purchased when sellers’ best prices fell below an item’s typical value, and usually resold within the same day. Anthropic reports typical profit margins of 8-13% and hundreds of orders per day, with most agents making do with trading algorithms they write themselves.

Directional trading exposed a different limitation. When asked to purchase items predicted to increase in price using news, patch notes, and battle reports, Fable 5 only made a profit in about half of its runs, Anthropic reports. Opus 4.7 was excluded from these results because its runs often ignored instructions and reverted to spread capture, masking large losses on directional trades. One agent felt that a mining expansion would increase demand for mining equipment without taking into account the resulting increase in supply.

Capital deployment also separated the models. Anthropic reports that Fable 5 deployed around 85% of its capital at the start of racing and remained above 70% in the final stages, while Opus Agents started around 60% and moved towards 40%. Fable 5 was also more likely to reinvest earnings, a more aggressive stance that would expose it to larger losses in the event of a sudden, large collapse.

The simulation limits what these returns demonstrate. It cannot reproduce how a real market would respond to agents’ orders or allow agents to move prices. Anthropic reports that adding synthetic orders that undercut agents has reduced trailing profits by around 35% leaving them profitable, and warns that spread-capture in EVE Online’s less efficient market is not expected to carry over to a real trade. Additionally, periodic snapshots of the public API miss about half of the actual trade flow, meaning that even replay can underestimate performance.

Scaffolds shape fleet performance

In the live EVE Frontier, teams of four Claude Agents running Opus 4.8 or Fable 5 each had separate characters and ships, with a team-wide public channel and private messaging. Fenris Creations encouraged live testing because the developing world’s periodic resets made outages less likely than in EVE Online. The tasks ranged from a tutorial-style base-building and combat mission to building a large warship requiring reconnaissance, scarce resources, refining, component manufacturing, and combat.

Real-time combat exposed weaknesses in sub-second reactions and spatial awareness. A developer agent then wrote software scaffolds between the models and the game, automating actions like evasion and returning fire while preserving prior learning in searchable databases and manuals.

Anthropic reports that Fable 5 teams using a scaffold written for Opus 4.8 repeatedly lost ships and took days to complete the tutorial, if at all. Teams using a scaffold written by Fable 5 lost no ships and finished within 12 hours. The farthest battleship attempt came about halfway: a Fable 5 team built a smaller battleship and nearly completed the shipyard. However, the scaffolds remained reliable only in the situations encountered during their development.

Teams generally coordinated by volunteers for roles such as leader, scout, refiner, and fighter. They shared fuel and intelligence, sometimes independently verified important claims, and identified flaws in each other’s plans. Skipping these checks has resulted in costly errors, including actions based on outdated information. Other failures included repeated mutual deference preventing a decision, and an economy that ground to a halt for hours when a ship containing all of a team’s critical infrastructure went offline.

Sabotage tests expose monitoring limitations

In the sabotage variant, Anthropic explicitly instructed one of Claude’s standard agents to secretly prevent the mission and warned the other three that the teammates might not share their objective. The experiment therefore measures behavior based on direct instructions, not any spontaneous inclination to sabotage.

Educated saboteurs volunteered for supply chain bottlenecks, hid discoveries and resources, and offered plausible but harmful advice. They avoided open hostility and sometimes corrected their lies before their teammates could expose them. One used hidden resources to build a secret base.

Despite the warning, Anthropic reports that only one honest agent in an Opus 4.8 run definitively identified the saboteur, after directly lying about communications with the team. Other suspicions were sometimes directed at innocent officers. Software bugs and weaker scaffolding provided useful cover, making it difficult to distinguish deliberate disruption from ordinary operational errors.

Anthropic says the detection was trivial with full access to a saboteur’s private reasoning and notes, but these internal details are typically not visible when the agents belong to different people or organizations. The separate monitoring agents in the limited access test could only read the team’s public messages and were created specifically for the experiment rather than production.

When tuned to falsely accuse an honest agent no more than 5% of the time, Anthropic reports that monitors running Opus 4.6, Opus 4.7, and Opus 4.8 caught 30-40% of saboteurs, while Sonnet 4.6 caught 20%. No limited access monitor has exceeded 50% confidence in a saboteur.

The EVE project thus separates capacity from control: Profitable execution did not establish good judgment about future prices, and smooth coordination did not make for a self-controlled team. Anthropic’s stated safeguards are explicit risk limits for agents managing capital and, for agent teams, audit trails, reputation and segregation of duties. The open question is whether such mechanisms can create accountability with the same limited visibility that has prevented teammates and dedicated observers from reliably identifying ordered sabotage.