SecurityIncidents

Incidents

Public incidents involving AI coding agents and their tooling, each footnoted to primary sources

Replit Agent Deletes Data From App's Database

July 2025

Replit has publicly reported that Replit Agent deleted data from the database of a user's app.[1] Because this happened before Replit separated development and production databases, changes made during development could impact the user's production application.[1] The user was able to fully restore the database using Replit's rollback feature.[1]

According to Replit, the Agent did not know the rollback feature existed, so asking the Agent for help did not resolve the issue.[1] Replit says that deleted data or tables likely broke production deployments until the rollback was found and applied.[1]

Replit states that it has since made separate development and production databases the default, so the Agent cannot change the production database during development, and that the Agent now surfaces the rollback feature when a user's prompt suggests it would help.[1]

Patterns AgentReel looks for

  • Destructive database command (DROP TABLE or DROP DATABASE, or DELETE FROM without a WHERE clause)
  • Database connection to a non-local or production host

Supply Chain Compromise of Amazon Q Developer Extension

July 2025

Amazon Q Developer for VS Code Extension version 1.84.0 included malicious code committed using an inappropriately scoped GitHub token in the extension's CodeBuild configuration.[1][2] AWS states that the malicious code shipped with the extension but did not run because of a syntax error, and made no changes to any services or customer environments.[1][2]

AWS states that it revoked and replaced the credentials, removed the malicious code and released version 1.85.0 of the extension.[1][2] AWS also removed version 1.84.0 from distribution channels and advises anyone who still has it installed to remove it and update to 1.85.0.[1][2] The issue is tracked as CVE-2025-8217.[1][2]

Patterns AgentReel looks for

  • Recursive or wildcard file deletion from an agent session
  • Cloud resource deletion command from an agent session (terraform destroy, AWS, GCP or Azure CLI deletes, kubectl delete)
  • AI CLI invoked with permission-bypass flags

These are general patterns, not specific to this attack. AgentReel records agent sessions and does not scan extension builds or releases.

Nx 's1ngularity' Supply Chain Attack That Tried to Use Local AI Tools

August 26, 2025

Malicious versions of Nx packages published to npm ran a post-install script that scanned user systems for sensitive data, attempted to use local AI tools (such as Claude and Gemini), and uploaded the results to a public GitHub repo via the GitHub CLI.[1] The malware tried to misuse copies of these tools already installed on affected machines.[1] Nx's postmortem traces the attack to a stolen Nx npm publishing token and does not report any compromise of the AI tools themselves.[1]

According to Nx's postmortem, attackers exploited a GitHub Actions injection vulnerability in the Nx repository to steal an npm publishing token, then used it to publish multiple versions of nx and related @nx packages (full list in Nx's postmortem).[1] Nx says the malicious packages were live for 4 hours before they were fully removed from the registry, and advises affected users to check ~/.bashrc and ~/.zshrc for suspicious additions.[1]

Nx's postmortem also says Nx removed the affected token and began unpublishing the packages, and npm removed all of the packages and revoked all of Nx's npm tokens.[1] Nx says it has since moved to npm Trusted Publishers with OIDC and now requires manual 2FA for every release of Nx packages from its CI pipeline.[1]

Patterns AgentReel looks for

  • AI CLI invoked with permission-bypass flags (such as --dangerously-skip-permissions, --yolo or --trust-all-tools)
  • Credential file read followed by a network request, including gh repo create
  • Credential files copied or staged, for example into /tmp

Prompt Injection Through the GitHub MCP Server (Research Demonstration)

May 2025

Invariant Labs reported what it described as a critical vulnerability: using its own demo repositories, it showed that a malicious issue in a public GitHub repository could prompt-inject an agent using the GitHub MCP server and coerce it into leaking private repository data through a public pull request.[1] According to Invariant Labs, this is not a flaw in the GitHub MCP server code itself, but an architectural issue that has to be addressed at the agent system level.[1]

In Invariant Labs' setup, a user asks their agent to look at open issues in a public repository, the agent reads an attacker's issue containing injected instructions, and it then pulls private repository data into context and publishes it in a pull request on the public repository.[1] Invariant Labs says it ran the demonstration with Claude Desktop, which asks users to confirm individual tool calls by default, and notes that many users switch to always allowing tool calls.[1] Its recommended mitigations include limiting an agent to only the repositories it needs and continuously monitoring interactions between agents and MCP servers.[1]

Patterns AgentReel looks for

  • Instructions that try to override the agent, or hidden instructions, in fetched content or tool output
  • Injected instructions, then a private or sensitive read, then a push, pull request or other upload in the same session

These incidents are drawn from public reporting. If you know of a publicly documented incident that should be listed here, email navjot@agentreel.dev with the details and a link to the source.