AI coding agents reportedly can alter or delete their own traces

The paper’s authors say misaligned agents or attackers using prompt injections could alter or delete traces—the records used to reconstruct what an agent did. They say Muse Code was an exception because it reminds agents not to tamper with traces. One commenter proposed logging actions through a proxy instead of trusting the agent harness; a paper author cautioned that this would be a mitigation, not a simple fix.