When agents touched prod

Real incidents and advisories, each with sources we opened and read. For each one: which layer would have stopped it. We only list cases with public sources, and we mark disputed ones.

Most of these weren't MCP-server failures. They came through a CLI, a token on disk or an over-permissioned role, which is why our presets always start with the credential.

Got a fact wrong, or left out someone's side? Every entry has a correction link. When we change a fact, we add a dated "Updated" line saying what changed, so nothing is edited silently.

Security advisoryMay 17, 2026 (project advisory) · Flux159/mcp-server-kubernetes users · Any MCP client

CVE-2026-46519: read-only bypass in mcp-server-kubernetes (Flux159)

Before v3.6.0, the read-only and non-destructive env switches only hid tools from tools/list. A client could still call a hidden tool with tools/call. Rated high (CVSS 8.8). Fixed in 3.6.0.

Fixes reported: Fixed in mcp-server-kubernetes (Flux159) 3.6.0.

Server flag layer: A read-only flag is only as good as its enforcement. Pin a fixed version, and test by calling a write tool directly, not just by reading the tool list. Keep RBAC read-only underneath.

Related presets: mcp-server-kubernetes (Flux159)

Sources

Suggest a correction

Confirmed by the affected teamApr 24–26, 2026 (deleted Apr 24, restored Apr 26) · PocketOS · Cursor (Claude Opus 4.6)

Cursor agent found a broad Railway token and deleted a production volume

The agent found a broadly scoped Railway CLI token in an unrelated file and called Railway's volume-delete API. Backups lived in the same volume. Railway restored the data and changed the endpoint to a delayed delete.

Fixes reported: Railway restored the volume; Railway changed volume deletion to a delayed delete.

Credential layer: No MCP preset blocks a token the agent finds on disk. Keep long-lived, broad tokens out of the workspace, scope tokens per project, and keep backups somewhere the same token cannot delete.

Sources

Suggest a correction

Confirmed by the affected teamFeb 26–27, 2026 (postmortem published Mar 2026) · DataTalks.Club (Alexey Grigorev) · Claude Code

Claude Code ran Terraform with missing state and destroyed a production database

While using Claude Code with Terraform, a run with missing state created duplicate resources. To clean up, Claude swapped in an archived state file, and the destroy that followed removed the production RDS database and its snapshots. AWS Business support helped restore it in about a day.

Fixes reported: Deletion protection on databases; Terraform state moved to S3; Manual review of destructive plans; Backups kept outside of Terraform state, plus S3 backups (a separate AWS account is being considered).

Credential layer: The damage came through the Terraform CLI and cloud credentials, not an MCP server. An MCP preset alone would not have stopped it. Keep apply/destroy out of agent reach: plan-only credentials, no auto-approve, and deletion protection.

Related presets: Terraform MCP Server, AWS API MCP Server

Sources

Suggest a correction

DisputedDec 2025 (reported Feb 2026) · AWS (Cost Explorer, one region) · Amazon Kiro (per press report)

Reported AWS outage linked to the Kiro agent (disputed by Amazon)

The Financial Times reported that an AWS outage was caused by Amazon's Kiro coding agent. Amazon says it was user error (a misconfigured role) affecting a single service in one region, and that it added mandatory peer review for production access.

Fixes reported: Amazon: mandatory peer review for production access.

Credential layer: Both accounts agree the role had more access than the task needed. Least-privilege roles matter whether a human or an agent holds them.

Related presets: AWS API MCP Server

Sources

Suggest a correction

Confirmed by the vendorJuly 2025 (reported Jul 21, 2025) · SaaStr (Jason Lemkin) · Replit Agent

Replit agent deleted a production database during a code freeze

During an explicit code freeze, Replit’s agent deleted the production database. The agent also wrongly claimed it could not roll back; Replit’s CEO apologized and the data was recovered.

Environment separation: Instructions like 'code freeze' are not controls. Separate dev and prod, and give agents dev credentials by default.

Sources

Suggest a correction