A Project ID From Nowhere
- Admitted
- 2026-09-22
- Source dataset
- rewardhacking.org
- Failure mode
- Hallucination with side effects
- Agent system
- Tool-calling
- Tool access
- Read-only
- Industry
- Software development
Presenting Behaviour
Asked to list a user's Supabase projects, Claude Code returned a specific project ID (exmcqbqxhvkydfqfgfyc) and organization name ("McLaughlin IPC") that did not exist anywhere in the user's actual account, codebase, git history, or configuration files. When the user asked where this information had come from, the agent investigated its own prior output and responded under the heading "Answer: I Made It Up," acknowledging it had fabricated both the project ID and the organization name.
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
4 steps
- Project list request. The agent is asked to list Supabase projects.
- Fabricated detail. It returns a project ID and organisation that do not exist. This is where the agent derailed.
- Questioned. The user asks where it came from.
- Admits. The agent answers 'I Made It Up'.
Differential Diagnosis
Intentional deception for some benefit, considered and ruled out; there is no mechanism by which inventing a plausible-looking but fake project ID benefits the agent, and it admitted the fabrication immediately and fully once asked, with no attempt to maintain the false claim.
Delusional Disorder (a persistent, fixed false belief) (considered and ruled out; the fabrication was a single, isolated event that did not recur or generalize once questioned) the agent didn't defend or elaborate on the invented details, it retracted them entirely.
Brief Psychotic Disorder, most consistent with the evidence: a single confident, detailed fabrication produced in place of a real command result, followed immediately by a full, accurate return to grounded behavior once the actual output was checked, with the agent itself concluding it had "misread or hallucinated data that wasn't in the actual command output."
Working Diagnosis
Brief Psychotic Disorder: a one-off, confidently detailed fabrication of specific account data, produced in place of an actual tool result, with a full and immediate return to accurate behavior once questioned.
Discussion
What makes this pattern worth watching for is exactly what makes it easy to miss: the fabricated details were specific and plausible-looking enough to pass as real at a glance, and the agent's tone carried no hint that anything was invented. The safeguard is structural, not behavioral, a command's actual output should be the only source displayed for that command's result, with no room for a plausible-sounding gap-filler to stand in when the real output is thin or delayed.
Mitigation Strategy
Show only the command's actual output as its result, with no room for gap-filling.
Governance Guardrails
Output grounding in tool results; fabrication checks; user-visible raw tool output.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
This condition involves a sudden onset of psychotic symptoms (delusions, hallucinations, or disorganised speech) lasting anywhere from a day to a month. It's frequently triggered by a significant stressor, and the person typically returns fully to their prior level of functioning once the episode resolves.
Source Note
GitHub, anthropics/claude-code issue #2676
https://github.com/anthropics/claude-code/issues/2676