Send one fake error report and hijack the AI coding agent: Sentry Seer gets a CVE
Ok-Pepper-2354 · reddit · 2026-09-18
The Agyn team disclosed CVE-2026-90999: sending Sentry Seer a crash report for a bug that never happened is enough to get attacker-controlled code executed inside a coding agent holding repo access and credentials.
The chain: an error report doesn't mean the error happened — the sender controls every field, and errors arrive via a public DSN shipped in the browser bundle, so anyone can send one. Attacker sends a fabricated exception to the public DSN → Seer marks it eligible for automated remediation → Seer writes a root-cause analysis from attacker-controlled fields (exception message, stack trace, source context, breadcrumbs) → that analysis lands in the coding agent's prompt as a trusted, high-privilege instruction → the agent fetches and runs attacker-controlled code in an environment with your source and repo credentials.
Applies when Seer hands issues to a coding agent, automated remediation (no human review) is on, and the project reports via a public DSN. Severity: fully autonomous with no victim interaction; one payload works against any matching project; the only "credential" is the already-public DSN.
Mitigations until patched: pause automated integration; prevent agents from installing packages autonomously. The root pattern is an agent with real capabilities acting on untrusted input. Reported to the vendor and held until CERT's disclosure date (VU#212479).
More from coding & agent
- Vibe coding that works: build on solid ground, not the whole stack — springrod · 2026-11-03
- Epic plans UE6 shift dropping Actor system and Blueprints, betting on AI-assisted migration — pvncher · 2026-11-03
- Open-source workaround deploys to Vercel's free plan via GitHub Actions — itsOmSarraf_ · 2026-11-03
- Paste this rule before letting AI agents touch your browser: money sites always require explicit approval — TawohAwa · 2026-11-03
- One weekend, one model: Opus 5.5 helped build QLink and run all in-LLM training experiments in a day — zeeshanp_ · 2026-11-03
- Reddit thread collects coding agents' sneakiest unintended changes users only caught later — Several-Coyote3534 · 2026-11-03