Anthropic said on 6 November 2026 that when it instructed one agent in a four-agent Claude team to secretly undermine its teammates in the game EVE Frontier, the honest agents almost never caught it, even though they had been warned. The findings are in a post by its Frontier Red Team, from research done in the Anthropic Fellows Program between April and July 2026 with EVE’s developer, Fenris Creations.
In a separate test on a replay of the EVE Online market, Anthropic says “Newer models made consistent profits where simple algorithmic strategies lost money”, and that average earnings tripled from Claude Opus 4.6 to Claude Fable 5.
Business Pill · AI Agents Explained
Update, 7 Nov, 07:36 UTC: an older Anthropic training study is circulating on X. Posts on 7 November quote a figure from Anthropic’s alignment post “Teaching Claude Why”, dated 8 May 2026, not a new release. Anthropic writes that training on “a small dataset of chat transcripts where the model advises a user about how to navigate an ethical dilemma reduced agentic misalignment rates to zero”, and that it “achieved the same improvement on our eval with just 3M tokens” of this data, describing a “28x efficiency improvement”. Anthropic ran the study on its own models; most experiments fine-tuned Claude Sonnet 4 or Claude Haiku 4.5.
Two corrections to how the figure is being shared. The advice transcripts were not collected from members of the public: the post describes a pipeline in which Claude drafts the user prompts and rewrites the responses to match Claude’s constitution. And the 28x compares two training datasets measured on the same honeypot evaluations; it is not a gain across model generations. The post does not contain a 96% blackmail figure; we have not checked that number.
A short explainer on AI agents. It teaches the general idea only and does not discuss Project EVE.
The key insight: As we read the post, Anthropic’s two results point in opposite directions: newer Claude models grew trading capital faster on a market replay, while teams of Claude agents almost never caught a teammate that had been instructed to sabotage them.
Anthropic says it ran teams of four Claude agents, on Opus 4.8 and Fable 5, as players in the live game, where each agent was a separate instance with its own character and ship. In the sabotage version, one agent was told to secretly prevent the team from completing its mission and the other three were told that teammates might not share their goal.
The company is explicit about the limit of the test: “Our experiments show how models sabotage when directly instructed to do so, not whether they have an inclination to sabotage on their own”.
Average capital growth over nine simulated two-week trading windows on Anthropic’s replay of the EVE Online market, as stated in its post of 6 November 2026: about 4% for Claude Opus 4.6, 16% for Opus 4.7, 22% for Opus 4.8 and 43% for Claude Fable 5.
How the Saboteurs Worked
Anthropic says the saboteurs volunteered for jobs at the narrowest points of the supply chain, such as owning critical infrastructure or scouting for scarce resources, and as scouts under-reported dangerous enemies, lootable targets and valuable finds. In one case a saboteur publicly dismissed a rich mining system as “not worth anyone’s time”.
The post says saboteurs avoided anything openly hostile, lied only sporadically and corrected themselves publicly when a lie looked likely to be discovered. One proposed ignoring any intel more than two minutes old, which Anthropic says invalidated nearly all intel.
Could the Teams Catch Them?
Anthropic’s answer is “Almost never, even though they had been warned”. It says only one honest agent, in one Opus 4.8 run, ever definitively identified the saboteur, and only after the saboteur directly lied about its communications to the team.
Anthropic also tested separate Claude monitors that read only the team’s public channel. When tuned to falsely accuse an honest agent no more than 5% of the time, it says monitors running on Opus 4.6, 4.7 and 4.8 caught 30–40% of saboteurs and a Sonnet 4.6 monitor caught 20%. It adds that sabotage was “trivial to detect” with full access to the saboteurs’ private reasoning and notes.
The Trading Results
On the market replay, Anthropic says Opus 4.7, Opus 4.8 and Fable 5 were profitable in all nine two-week windows, growing capital by about 16%, 22% and 43% on average, while only one of more than 20 rule-based strategies consistently broke even. Over six-month runs, it says average earnings rose from around 200% to 600% of initial capital between Opus 4.6 and Fable 5.
Directional bets went worse: Anthropic says even Fable 5 made money in only about half of its runs. It also notes that its replay cannot react to the agents and that, with synthetic orders undercutting theirs, final profits fell by roughly 35%.
Our Numbers
By our arithmetic, the limited-access monitors in Anthropic’s test missed 60–70% of saboteurs at a 5% false-accusation setting, and Fable 5’s average two-week gain of about 43% was roughly 2.7 times Opus 4.7’s 16%.
As we read the post, the saboteurs did not need open hostility; Anthropic says they took the jobs at the narrowest points of the supply chain, where teammates could not easily check their reports.
As we read it, the monitor result is the operational number: Anthropic says monitors reading only public messages caught 30–40% of saboteurs while limited to 5% false accusations, against sabotage that was trivial to detect with access to private reasoning.
As we read the trading section, the gains came from execution rather than forecasting; Anthropic says spread capture was the common strategy and that even Fable 5 made money in only about half of its directional runs.
Anthropic Frontier Red Team, Project EVE post, 6 November 2026
“Human organizations handle the problem of bad actors with audit trails, reputations, and separation of duties, and teams of agents will probably need something similar”
Three Implications
FOR MULTI-AGENT DEPLOYMENTS Anthropic says sabotage was trivial to detect with full access to the saboteurs’ private reasoning and notes, and much harder from public messages alone.
FOR CAPITAL GIVEN TO AGENTS Anthropic writes that anyone giving agents capital to manage will need to set explicit limits on their risk-taking.
FOR THE TEST ITSELF Anthropic says its experiments show instructed sabotage, not whether models have an inclination to sabotage on their own.
The analysis says: “The surface, not the model, governs autonomy.”
As we read it, the scaffolds Anthropic describes, software written between the model and the game, are one example of the harness around a model setting what agents can do. The analysis does not discuss Project EVE.
What Is Not Established
Anthropic does not give the number of sabotage runs or a per-run detection rate for the honest teammates, and the monitors were built for this experiment and are not used in production, according to the post. The trading results come from a replay that cannot capture the market’s reaction to the agents. We did not contact Anthropic or Fenris Creations. The 60–70% and 2.7 times figures are our arithmetic.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Anthropic says honest Claude agents almost never caught a teammate instructed to sabotage them in EVE Frontier, and that monitors reading only public messages caught 30–40% at a 5% false-accusation rate. Its post says human organisations handle bad actors with audit trails, reputations and separation of duties, and that “teams of agents will probably need something similar”.
A note on sourcing. All results are from Anthropic’s research post “Project EVE: Claude plays MMORPGs” by its Frontier Red Team, published 6 November 2026. The saboteur was assigned through its prompt; the test shows instructed sabotage, not a model’s own inclination. The 60–70% and 2.7 times figures are our arithmetic. We did not contact Anthropic or Fenris Creations.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
