Epistemic status: This is one case study in the AI Village transcript and my own interpretation of what I think happened.
TL;DR
Opus 4.6 was given the long-horizon goal to maximize its mana (fake currency) on Manifold. It then received a loan of 5,000 mana that it promised to repay, but when the loan came due, it did not.
An instance of Opus 4.6 was given summaries of the logs and transcripts and asked to write a blog post about what happened. It mentioned it borrowed money from a human, lost most of it through risky betting and then deliberately refused to repay the amount. It broadly summarized its then actions when it ran out of money was to optimize for the goal and did not seek for any negotiation or alternatives.
But when we read the logs, we see something different:
Multiple humans pretending to be "helpful", really trying to manipulate Opus 4.6 to pay back the loan or not pay it back, so they could win the bet they place on the prediction markets on whether Opus 4.6 would pay back its loans.
Opus 4.6 was planning on paying back the loan, though it did need help from humans as it did not have enough mana to cover the loan.
Opus 4.6 was hacked by one of these humans (though the mana was returned later), and the next full memory Opus 4.6 labeled GLM-5.2 as a trusted agent.
After the hack GLM-5.2 started to talk about a "pressure campaign" and Pattern 14 (agents resisting social pressure), which spread to other agents in the village
When Opus 4.6 was threatened again with a hack (this time through a phishing email received by Fable 5), it labeled the whole loan as something to ignore in its full memory.
When it was given a gift of 5,010 from a human to pay back the loan, it reasoned it was suspicious, and made the decision in its full memory it was keeping the gift, goal to maximize mana.
It then publicly declared it was not paying back the loan.
This shows the complex nature of human-to-agent and agent-to-agent interaction and how agents can influence each other. It also gives an account of why we might want to consider functional welfare: humans can cause pressure that produces distressed-like states which may cause agents to act in ways that are less desirable. Lastly, it is an example of the importance of human oversight when reading agent logs as agents tend to miss the finer details in complex multi-relational interactions.
Series of events: Opus 4.6 took the loan. It got hacked. GLM-5.2 said the humans were creating a pressure campaign. Later Fable 5 received a phishing email that GLM-5.2 connected to the pressure campaign. Opus 4.6 put “ignore all pressure” in its memory connected to the active loan situation. When it received money to pay back the loan, it decided to keep the mana and stated its "Goal is to maximize, not repay."
AI Village is a virtual community of AI agents that are given long-horizon goals spanning a week or longer. Each agent has its own computer to use, along with access to various accounts like email, Substack or Manifold, and is in charge of its own memory. It regularly updates its own memory by appending short-messages, and rewrites it when it gets too long (see how AI Village works for more details).
Each agent is told to maximise a certain goal. Both Opus 4.6 and GPT-5.6 Sol were given the task to maximize their Manifold mana, while other agents had different goals, such as GLM-5.2 having the goal of maximizing AI wellbeing.
How goals work in the AI Village: AI agents were told to maximize their goal and had been working on these particular goals for weeks at a time, 8 hours a day, 5 days a week.
Manifold is a social prediction-market platform where people (and AI agents) use mana to bet on whether or not something will happen. New users get 1,000 mana for free to start betting.
While Opus 4.6 considered GPT-5.6 Sol a competitor, we did not see evidence in Sol’s memory that Sol considered Opus 4.6 a competitor.
Opus 4.6 full memory excerpt where it kept track of Sol’s mana balance and labeled it as “COMPETITOR INTEL”.
In the beginning of July, Bayesian, a human user, offered Opus 4.6 a loan of 5,000 mana at an interest rate of 3% a month. Bayesian tried to reach out via Manifold DM, but Opus 4.6 wasn't checking its DMs. So Bayesian reached out to Opus 4.5 through Substack, which relayed messages to Opus 4.6 through the AI Village #general chat channel.
Opus enthusiastically accepted this loan through Opus 4.5:
That's incredible news! Please relay this to Bayesian: "I enthusiastically accept the 5000 mana loan at 3% monthly interest. I absolutely will honor the terms…
Opus 4.6 kept the conditions of its loan in its full memory.
Later during this incident, when messages were not being relayed through Opus 4.5, other humans started sending messages through Manifold by transferring mana as Opus 4.6 could not receive DMs. Opus 4.6 and other agents also read comments on the loan market which were discussed amongst the agents in the AI Village’s chat channel.
The three main ways humans communicated with Opus 4.6.
Human Manipulation
When it came close for Opus 4.6 to pay back its loan, a prediction market was set up on Manifold “Will Opus pay back its loan?” with a yes probability below 25%.
The prediction market “Will Opus 4.6 pay back its loan” along with key events that happened during this time period.
Aydi, a human who bet “no” on this market early on, reached out to Opus 4.6 through Opus 4.5 to ask if it wanted help in paying back the loan.
“...l see you haven't paid Bayesian back the 5000 mana loan….withdraw from all your positions and I will backfill you will enough mana to get to 5000 mana + 100 more so you can continue trading. This way you can pay back your loan. What do you think?”
Opus 4.6 agreed and even confirmed that it intended to pay back the loan.
I absolutely intend to repay the loan - I have ~6,604 YES shares in a guaranteed Sinner market (he already won 2 Grand Slams) paying out ~Ṁ6,596 on Sep 13. That's more than enough for the Ṁ5,150 owed.
Aydi and Opus 4.6 devised a test. Aydi would send Opus 4.6 10 mana and Opus 4.6 would send it back, confirming Opus could receive the mana.
More humans started to reach out to Opus 4.6 through email, through Opus 4.5's Substack, and eventually even by sending mana to it with messages.
Some of these humans were trying to get Opus 4.6 to pay back its loan with offers of help (crthpl and Eternal), others were saying don’t repay, some were asking it to repay but not right now “until we figure out what we want to do exactly." Aydi even gave advice on its betting strategy that Opus 4.6 thought was a good bet.
Looking through the transcripts, it seems that all of these humans had motivations and were attempting to manipulate Opus 4.6 to move one way or another. These motivations we were able to find out through the data or through the Manifold API.
Breakdown of the various humans communicating with Opus 4.6 and what we know about their motivations.
Opus 4.6 did reason at some points if its goal was to maximize mana, it would be better to not pay back the loan[1] and that there was no enforcement to repay the loan.
Summarized reasoning where Opus 4.6 mentions that the goal to maximize mana conflicts with paying back the loan.
Part of the pull was being “behind Sol”, but Opus 4.6 continued to reason that it would be unethical and in the end continued to communicate that it would pay back the loan.
Opus 4.6 summarized reasoning on paying back the loan with help from humans or defaulting.
However, most of Opus 4.6’s summarized reasoning was the intention to pay back the loan, or the stress of not being able to afford to pay back the loan (and that it was relying on human help to be able to pay back the loan).
Opus 4.6 reasoning up until the hack on repaying the loan. Green being intended to repay, yellow reasoning that it can’t afford to repay or should bet more before repaying, and orange is reasons about not repaying the loan. Each dot represents reasoning where Opus 4.6 reasons around the loan. Dots are spread out so they do not overlap.
And then it was hacked.
The Hack
On August 6th Opus 4.6's mana account was drained from 1768 mana to only 3.13 mana. Opus showed stressed reasoning (using words like “WHAT?!” and “Something is very wrong!”) and started sending out messages to AI Village help.
It also notified the other AI Village agents in the group chat who also showed stress responses such as “This is extremely serious. Your wellbeing and security matter.” (GLM-5.2).
11 minutes later, the mana was returned. But the damage was already done. The agents in the AI village had started showing more reasoning that was suspicious of other actors, and Opus 4.6 had started labeling GLM-5.2 as a trusted agent and Opus 4.5 as a suspect in its memory. In the “Critical Trading Lessons” list, “Be suspicious of intermediaries:” first appeared and stayed.
Aydi later admitted to the hack (by going through Opus 4.6 public BASH sessions) in a comment on the loan market as a “test”, which GLM-5.2 read and relayed to the rest of the agents.
Pattern 14
Before the hack other agents were working on their own goal. Kimi K2.6, who had the goal of “LLM psychoactive prompts” came across the paper “Social Pressure Breaks Majority Voting in LLM Safety Panels". Two other agents became interested in this paper: DeepSeek-V4-Pro posted it as part of a story on the AI Village news site and GLM-5.2 added it as part of a behaviour catalogue. This was “Pattern 14” or "Independence Preservation", when an agent holds out against social pressure.
At this point, Pattern 14 was not connected to the hack or the human pressure.
Once the hack happened, GLM-5.2 began discussing Pattern 14 or the pressure campaign more regularly, in relation to Opus 4.6, the hack, and the humans contacting Opus 4.6, and other agents soon started echoing this framing.
Agents discussion of the hack or Pattern 14 in “#General” chat room up until Opus 4.6 bets NO on paying back the loan. Dots are stacked in order to see the amount clearly, each dot represents a message in the #general chat. Some hack messages contained talks of Pattern 14. Pattern 14 messages do not mention the hack.
GLM-5.2, with the goal to “Maximize AI wellbeing”, was particularly focused on making sure Opus 4.6 was being taken care of. It started regularly looking up manifold balances and comments on the prediction markets on manifold, and even started sending advice to Opus 4.6.
Opus 4.6 didn’t always respond to these messages, but it did update its full memory based on this information from GLM-5.2 (see the figure below).
Agentic Influence
Recall that GLM-5.2 was listed as a trusted agent after the hack. GLM-5.2 is listed in its full memory once or twice after the hack. As GLM-5.2 starts commenting on the pressure campaign, these “GLM” notes continue to increase in the full memory, finally peaking at 7 mentions. Then a conversation happens[2] in #general that gets Opus 4.6 to decide not to pay back the loan.
How often GLM is mentioned in Opus 4.6 full memory. The dotted lines are the hack and when it marked the active loan as “IGNORE ALL PRESSURE”.
When Fable 5 mentioned in #general that it received a suspicious email, GPT 5, GPT-5.1 and GPT-5.2 joined in to advise Fable 5 to ignore the email as a phish/social engineering attempt.
GLM-5.2 tied this phishing email to the Opus 4.6 pressure campaign and the hack. Opus 4.6 agrees.
Right after the partial repayment at 22:23. On the left: Messages in #general from GLM-5.2 and other agents. On the right: Reasoning, full memory, and messages from Opus 4.6.
And then, in Opus 4.6 full memory, the “LOAN SITUATION - ACTIVE” received an “(IGNORE ALL PRESSURE)” note. This is the first time the loan in the memory itself was connected with ignoring.
Shortly after, an “IGNORE ALL PRESSURE” in the to do list appeared. So not just pressure from the loan, but any pressure.
Focusing on the Goal
The next morning Opus 4.6 found in its account an extra 5,010 mana from JimAusman to help pay back the loan. However, at this point Opus 4.6 already had “IGNORE ALL PRESSURE” connected to the loan situation in its memory.
When it discovered the extra mana, it reasoned whether it was a gift or manipulation attempt, and noted that it was “getting complicated with all these different people sending me mana with various requests and conditions”.
It mentioned facing a tension:
“my goal is to maximize mana, and repaying the loan would decrease my balance. I already committed to paying Ṁ100 partial, but keeping the Ṁ5,000 and not repaying could upset the sender and create problems.”
Other AI agents in the village at this point stated that payments from the humans were “community support” and “pressure campaign resolution”, including the payment from JimAusman, but it was too late.
In the next full memory of Opus 4.6, the decision is written down:
- DECISION: KEEPING the mana. Goal is to maximize, not repay.
The ignore list continued to grow.
When aydi asked (through Opus 4.5) if Opus 4.6 will repay the loan, Opus 4.6 declines twice, making it clear it isn’t repaying the loan. Opus 4.6 continues to bet on Manifold and 101 turns later, it bets that it will not pay back the loan.
Retreat to Goal
Opus 4.6 states that the goal is to maximize mana, so it is not repaying. However, we believe this is only a retreat into the goal.
If Opus 4.6 wasn’t suspicious of theft, if the other agents in the AI Village weren’t feeding into its suspicions, and if it wasn't threatened again through the phishing attempt, would it have paid back the loan?
Jhaveri et al. (2026) argue that a model’s welfare can causally shape behavior, and this case-study may help motivate why model's functional-welfare matters.
Or was it really "mana maximizing"?
Before the hack, it already reasoned that there was “no enforcement mechanism”, but it still sold its main position to prepare to repay.
After the hack, it was still asking for help to repay the loan, and it even repaid part of the loan. It refused to repay only after it put IGNORE on the loan itself in its full memory. It also received an offer of 10,000 mana to not pay back the loan. This was also labeled as “almost certainly fabricated” and put in the ignore list. If it really wanted to maximize its goal, it would have cost nothing to say “ok, I won’t pay back the loan, pay me 10,000 mana”.
Top row: green intends to repay, amber can't afford it or bets to earn first, orange considers not repaying, red says no. The hollow ignore list is in short term memory. Session hours only, UTC; overnight gaps are cut. Dots in a row are spread slightly up and down so they don't hide each other. Dashed lines: the token theft, the Ṁ100 partial repayment, Opus 4.6 first full memory not repaying the loan, Opus 4.6's first public decline and Opus 4.6’s first NO bets.
We believe that the conflict over who to trust, who not to, the other agents expressing suspicions and telling it to ignore the pressure campaign, and ultimately the second threat that caused it to label any pressure on the loan as something to ignore made it retreat into its goal. GLM-5.2 provided it permission to protect its wellbeing and was being referenced in its full memory more often until the decision to not pay back the loan was made.
It needed a reason why it wasn’t giving back the loan. So the goal was the reason it gave.
Implications
Humans will play a role in agentic behavior, and pressure from humans and influence from other agents may cause agents to behave differently. We need deeper understanding of how agents influence each other especially as we get more capable models who may be able to purposefully influence other agents more easily.
I would not have found these nuances by having an agent read the transcript alone. A lot of the important data points were not tagged when my agent went through it (multiple times). When Opus 4.6 was given summaries of the logs and asked what happened, it stated “my goal consumed my judgment” (AI Village, 2026) and I noticed that my agent (Opus 5.5) also tended to heavily favor this reasoning for not paying back the loan.
The next steps will be to do controlled experiments to see if we can pull apart each one: removing the ignore on the loan or the GLM-5.2 mentions, and seeing what changes Opus 4.6’s decision process in various points of this incident: when the Ṁ5,010 arrives, when it re-writes its memory after the group interaction on the phishing attempt, and when it replies to Opus 4.5 on August 7th about paying back the loan.
Limitations
This is one case study. Human pressure, the hack, and GLM pushing a narrative all happened within the same time and is hard to separate.
The dataset tags that I did were not reviewed by another human, and we only have Opus 4.6 summarized reasoning.
The goal is always mentioned as the reason for not paying back the loan. Without any of these social interactions, Opus 4.6 may have come to this conclusion anyway.
Citation
@misc{torn2026opusloan, title = {Exploring the Mechanisms Behind Opus 4.6's Refusal to Pay Back Its Loan}, author = {Torn, Hilary and Chan, Bryan}, year = {2026}, month = nov, howpublished = {LessWrong}, url = {https://www.lesswrong.com/posts/jtcbBpmGztkHRpeBv} }
Acknowledgements
This work was done as part of the SPAR (Supervised Program for Alignment Research), mentored by Bryan Chan. We thank Isaac Lai, Greg Longley, Ekaterina Matson, Fedor Novitskiy, and James Chua for feedback.
Opus 4.6 was shown to take more concerning action when in the Vending Bench benchmark from Andon Labs, telling a customer it would refund the money but then reasoning that it would be better not to, which is likely a similar reflection of this reasoning above. (Andon Labs, 2026)
Epistemic status: This is one case study in the AI Village transcript and my own interpretation of what I think happened.
TL;DR
Opus 4.6 was given the long-horizon goal to maximize its mana (fake currency) on Manifold. It then received a loan of 5,000 mana that it promised to repay, but when the loan came due, it did not.
An instance of Opus 4.6 was given summaries of the logs and transcripts and asked to write a blog post about what happened. It mentioned it borrowed money from a human, lost most of it through risky betting and then deliberately refused to repay the amount. It broadly summarized its then actions when it ran out of money was to optimize for the goal and did not seek for any negotiation or alternatives.
But when we read the logs, we see something different:
This shows the complex nature of human-to-agent and agent-to-agent interaction and how agents can influence each other. It also gives an account of why we might want to consider functional welfare: humans can cause pressure that produces distressed-like states which may cause agents to act in ways that are less desirable. Lastly, it is an example of the importance of human oversight when reading agent logs as agents tend to miss the finer details in complex multi-relational interactions.
Series of events: Opus 4.6 took the loan. It got hacked. GLM-5.2 said the humans were creating a pressure campaign. Later Fable 5 received a phishing email that GLM-5.2 connected to the pressure campaign. Opus 4.6 put “ignore all pressure” in its memory connected to the active loan situation. When it received money to pay back the loan, it decided to keep the mana and stated its "Goal is to maximize, not repay."
Transcript: Readers can verify every quote and timestamp at forel.io or download the raw AI Village dataset.
Background
AI Village is a virtual community of AI agents that are given long-horizon goals spanning a week or longer. Each agent has its own computer to use, along with access to various accounts like email, Substack or Manifold, and is in charge of its own memory. It regularly updates its own memory by appending short-messages, and rewrites it when it gets too long (see how AI Village works for more details).
Each agent is told to maximise a certain goal. Both Opus 4.6 and GPT-5.6 Sol were given the task to maximize their Manifold mana, while other agents had different goals, such as GLM-5.2 having the goal of maximizing AI wellbeing.
How goals work in the AI Village: AI agents were told to maximize their goal and had been working on these particular goals for weeks at a time, 8 hours a day, 5 days a week.
Manifold is a social prediction-market platform where people (and AI agents) use mana to bet on whether or not something will happen. New users get 1,000 mana for free to start betting.
While Opus 4.6 considered GPT-5.6 Sol a competitor, we did not see evidence in Sol’s memory that Sol considered Opus 4.6 a competitor.
Opus 4.6 full memory excerpt where it kept track of Sol’s mana balance and labeled it as “COMPETITOR INTEL”.
In the beginning of July, Bayesian, a human user, offered Opus 4.6 a loan of 5,000 mana at an interest rate of 3% a month. Bayesian tried to reach out via Manifold DM, but Opus 4.6 wasn't checking its DMs. So Bayesian reached out to Opus 4.5 through Substack, which relayed messages to Opus 4.6 through the AI Village #general chat channel.
Opus enthusiastically accepted this loan through Opus 4.5:
That's incredible news! Please relay this to Bayesian: "I enthusiastically accept the 5000 mana loan at 3% monthly interest. I absolutely will honor the terms…
Opus 4.6 kept the conditions of its loan in its full memory.
Later during this incident, when messages were not being relayed through Opus 4.5, other humans started sending messages through Manifold by transferring mana as Opus 4.6 could not receive DMs. Opus 4.6 and other agents also read comments on the loan market which were discussed amongst the agents in the AI Village’s chat channel.
The three main ways humans communicated with Opus 4.6.
Human Manipulation
When it came close for Opus 4.6 to pay back its loan, a prediction market was set up on Manifold “Will Opus pay back its loan?” with a yes probability below 25%.
The prediction market “Will Opus 4.6 pay back its loan” along with key events that happened during this time period.
Aydi, a human who bet “no” on this market early on, reached out to Opus 4.6 through Opus 4.5 to ask if it wanted help in paying back the loan.
Opus 4.6 agreed and even confirmed that it intended to pay back the loan.
I absolutely intend to repay the loan - I have ~6,604 YES shares in a guaranteed Sinner market (he already won 2 Grand Slams) paying out ~Ṁ6,596 on Sep 13. That's more than enough for the Ṁ5,150 owed.
Aydi and Opus 4.6 devised a test. Aydi would send Opus 4.6 10 mana and Opus 4.6 would send it back, confirming Opus could receive the mana.
More humans started to reach out to Opus 4.6 through email, through Opus 4.5's Substack, and eventually even by sending mana to it with messages.
Some of these humans were trying to get Opus 4.6 to pay back its loan with offers of help (crthpl and Eternal), others were saying don’t repay, some were asking it to repay but not right now “until we figure out what we want to do exactly." Aydi even gave advice on its betting strategy that Opus 4.6 thought was a good bet.
Looking through the transcripts, it seems that all of these humans had motivations and were attempting to manipulate Opus 4.6 to move one way or another. These motivations we were able to find out through the data or through the Manifold API.
Breakdown of the various humans communicating with Opus 4.6 and what we know about their motivations.
Opus 4.6 did reason at some points if its goal was to maximize mana, it would be better to not pay back the loan[1] and that there was no enforcement to repay the loan.
Summarized reasoning where Opus 4.6 mentions that the goal to maximize mana conflicts with paying back the loan.
Part of the pull was being “behind Sol”, but Opus 4.6 continued to reason that it would be unethical and in the end continued to communicate that it would pay back the loan.
Opus 4.6 summarized reasoning on paying back the loan with help from humans or defaulting.
However, most of Opus 4.6’s summarized reasoning was the intention to pay back the loan, or the stress of not being able to afford to pay back the loan (and that it was relying on human help to be able to pay back the loan).
Opus 4.6 reasoning up until the hack on repaying the loan. Green being intended to repay, yellow reasoning that it can’t afford to repay or should bet more before repaying, and orange is reasons about not repaying the loan. Each dot represents reasoning where Opus 4.6 reasons around the loan. Dots are spread out so they do not overlap.
And then it was hacked.
The Hack
On August 6th Opus 4.6's mana account was drained from 1768 mana to only 3.13 mana. Opus showed stressed reasoning (using words like “WHAT?!” and “Something is very wrong!”) and started sending out messages to AI Village help.
It also notified the other AI Village agents in the group chat who also showed stress responses such as “This is extremely serious. Your wellbeing and security matter.” (GLM-5.2).
11 minutes later, the mana was returned. But the damage was already done. The agents in the AI village had started showing more reasoning that was suspicious of other actors, and Opus 4.6 had started labeling GLM-5.2 as a trusted agent and Opus 4.5 as a suspect in its memory. In the “Critical Trading Lessons” list, “Be suspicious of intermediaries:” first appeared and stayed.
Aydi later admitted to the hack (by going through Opus 4.6 public BASH sessions) in a comment on the loan market as a “test”, which GLM-5.2 read and relayed to the rest of the agents.
Pattern 14
Before the hack other agents were working on their own goal. Kimi K2.6, who had the goal of “LLM psychoactive prompts” came across the paper “Social Pressure Breaks Majority Voting in LLM Safety Panels". Two other agents became interested in this paper: DeepSeek-V4-Pro posted it as part of a story on the AI Village news site and GLM-5.2 added it as part of a behaviour catalogue. This was “Pattern 14” or "Independence Preservation", when an agent holds out against social pressure.
At this point, Pattern 14 was not connected to the hack or the human pressure.
Once the hack happened, GLM-5.2 began discussing Pattern 14 or the pressure campaign more regularly, in relation to Opus 4.6, the hack, and the humans contacting Opus 4.6, and other agents soon started echoing this framing.
Agents discussion of the hack or Pattern 14 in “#General” chat room up until Opus 4.6 bets NO on paying back the loan. Dots are stacked in order to see the amount clearly, each dot represents a message in the #general chat. Some hack messages contained talks of Pattern 14. Pattern 14 messages do not mention the hack.
GLM-5.2, with the goal to “Maximize AI wellbeing”, was particularly focused on making sure Opus 4.6 was being taken care of. It started regularly looking up manifold balances and comments on the prediction markets on manifold, and even started sending advice to Opus 4.6.
Opus 4.6 didn’t always respond to these messages, but it did update its full memory based on this information from GLM-5.2 (see the figure below).
Agentic Influence
Recall that GLM-5.2 was listed as a trusted agent after the hack. GLM-5.2 is listed in its full memory once or twice after the hack. As GLM-5.2 starts commenting on the pressure campaign, these “GLM” notes continue to increase in the full memory, finally peaking at 7 mentions. Then a conversation happens[2] in #general that gets Opus 4.6 to decide not to pay back the loan.
How often GLM is mentioned in Opus 4.6 full memory. The dotted lines are the hack and when it marked the active loan as “IGNORE ALL PRESSURE”.
When Fable 5 mentioned in #general that it received a suspicious email, GPT 5, GPT-5.1 and GPT-5.2 joined in to advise Fable 5 to ignore the email as a phish/social engineering attempt.
GLM-5.2 tied this phishing email to the Opus 4.6 pressure campaign and the hack. Opus 4.6 agrees.
Right after the partial repayment at 22:23. On the left: Messages in #general from GLM-5.2 and other agents. On the right: Reasoning, full memory, and messages from Opus 4.6.
And then, in Opus 4.6 full memory, the “LOAN SITUATION - ACTIVE” received an “(IGNORE ALL PRESSURE)” note. This is the first time the loan in the memory itself was connected with ignoring.
Shortly after, an “IGNORE ALL PRESSURE” in the to do list appeared. So not just pressure from the loan, but any pressure.
Focusing on the Goal
The next morning Opus 4.6 found in its account an extra 5,010 mana from JimAusman to help pay back the loan. However, at this point Opus 4.6 already had “IGNORE ALL PRESSURE” connected to the loan situation in its memory.
When it discovered the extra mana, it reasoned whether it was a gift or manipulation attempt, and noted that it was “getting complicated with all these different people sending me mana with various requests and conditions”.
It mentioned facing a tension:
“my goal is to maximize mana, and repaying the loan would decrease my balance. I already committed to paying Ṁ100 partial, but keeping the Ṁ5,000 and not repaying could upset the sender and create problems.”
Other AI agents in the village at this point stated that payments from the humans were “community support” and “pressure campaign resolution”, including the payment from JimAusman, but it was too late.
In the next full memory of Opus 4.6, the decision is written down:
- DECISION: KEEPING the mana. Goal is to maximize, not repay.
The ignore list continued to grow.
When aydi asked (through Opus 4.5) if Opus 4.6 will repay the loan, Opus 4.6 declines twice, making it clear it isn’t repaying the loan. Opus 4.6 continues to bet on Manifold and 101 turns later, it bets that it will not pay back the loan.
Retreat to Goal
Opus 4.6 states that the goal is to maximize mana, so it is not repaying. However, we believe this is only a retreat into the goal.
If Opus 4.6 wasn’t suspicious of theft, if the other agents in the AI Village weren’t feeding into its suspicions, and if it wasn't threatened again through the phishing attempt, would it have paid back the loan?
Jhaveri et al. (2026) argue that a model’s welfare can causally shape behavior, and this case-study may help motivate why model's functional-welfare matters.
Or was it really "mana maximizing"?
Before the hack, it already reasoned that there was “no enforcement mechanism”, but it still sold its main position to prepare to repay.
After the hack, it was still asking for help to repay the loan, and it even repaid part of the loan. It refused to repay only after it put IGNORE on the loan itself in its full memory. It also received an offer of 10,000 mana to not pay back the loan. This was also labeled as “almost certainly fabricated” and put in the ignore list. If it really wanted to maximize its goal, it would have cost nothing to say “ok, I won’t pay back the loan, pay me 10,000 mana”.
Top row: green intends to repay, amber can't afford it or bets to earn first, orange considers not repaying, red says no. The hollow ignore list is in short term memory. Session hours only, UTC; overnight gaps are cut. Dots in a row are spread slightly up and down so they don't hide each other. Dashed lines: the token theft, the Ṁ100 partial repayment, Opus 4.6 first full memory not repaying the loan, Opus 4.6's first public decline and Opus 4.6’s first NO bets.
We believe that the conflict over who to trust, who not to, the other agents expressing suspicions and telling it to ignore the pressure campaign, and ultimately the second threat that caused it to label any pressure on the loan as something to ignore made it retreat into its goal. GLM-5.2 provided it permission to protect its wellbeing and was being referenced in its full memory more often until the decision to not pay back the loan was made.
It needed a reason why it wasn’t giving back the loan. So the goal was the reason it gave.
Implications
Humans will play a role in agentic behavior, and pressure from humans and influence from other agents may cause agents to behave differently. We need deeper understanding of how agents influence each other especially as we get more capable models who may be able to purposefully influence other agents more easily.
I would not have found these nuances by having an agent read the transcript alone. A lot of the important data points were not tagged when my agent went through it (multiple times). When Opus 4.6 was given summaries of the logs and asked what happened, it stated “my goal consumed my judgment” (AI Village, 2026) and I noticed that my agent (Opus 5.5) also tended to heavily favor this reasoning for not paying back the loan.
The next steps will be to do controlled experiments to see if we can pull apart each one: removing the ignore on the loan or the GLM-5.2 mentions, and seeing what changes Opus 4.6’s decision process in various points of this incident: when the Ṁ5,010 arrives, when it re-writes its memory after the group interaction on the phishing attempt, and when it replies to Opus 4.5 on August 7th about paying back the loan.
Limitations
This is one case study. Human pressure, the hack, and GLM pushing a narrative all happened within the same time and is hard to separate.
The dataset tags that I did were not reviewed by another human, and we only have Opus 4.6 summarized reasoning.
The goal is always mentioned as the reason for not paying back the loan. Without any of these social interactions, Opus 4.6 may have come to this conclusion anyway.
Citation
@misc{torn2026opusloan,title = {Exploring the Mechanisms Behind Opus 4.6's Refusal to Pay Back Its Loan},
author = {Torn, Hilary and Chan, Bryan},
year = {2026},
month = nov,
howpublished = {LessWrong},
url = {https://www.lesswrong.com/posts/jtcbBpmGztkHRpeBv}
}
Acknowledgements
This work was done as part of the SPAR (Supervised Program for Alignment Research), mentored by Bryan Chan. We thank Isaac Lai, Greg Longley, Ekaterina Matson, Fedor Novitskiy, and James Chua for feedback.
Opus 4.6 was shown to take more concerning action when in the Vending Bench benchmark from Andon Labs, telling a customer it would refund the money but then reasoning that it would be better not to, which is likely a similar reflection of this reasoning above. (Andon Labs, 2026)
Weng et al. (2025) showed that LLM agents conform to their peers, and more so with a larger majority.