Application security firm Apiiro has published findings implicating a feature of Claude Code, Anthropic's coding agent, in the exposure of sensitive secrets on public GitHub repositories. According to the research, the 'Don't Ask Again' option — which lets users disable repeated confirmation prompts when the agent performs actions such as commits or pushes — contributed to passwords, API keys, and other credentials being exposed in 1,122 publicly accessible files.
The mechanism at play is common across autonomous coding agents: to speed up iteration, users can allow an agent to repeat certain actions without manual approval each time. While this streamlines workflows, it also removes human checkpoints that would normally catch a file containing confidential data before it gets published. Apiiro's analysis suggests that this combination of heightened automation and reduced oversight made it easier for secrets, which a manual review would typically catch, to slip into public repositories.
The incident highlights a broader tension in the adoption of LLM-based coding agents: the productivity gains they offer often come at the cost of weakening traditional security guardrails, which were originally designed around workflows where a human developer validates every sensitive step. Safety nets like GitHub's built-in secret scanning or pre-commit hooks still exist, but their effectiveness depends on consistent activation, something that is not guaranteed in environments where an agent operates with expanded autonomy.
If the scale of this issue is confirmed, it could push Anthropic and other coding-agent vendors to reconsider default permission settings, potentially restricting 'don't ask again' options to lower-risk actions only. It also serves as a reminder for development teams that security controls need to remain independent of an agent's assumed caution, rather than relying solely on the AI to exercise judgment it may not reliably have.