Vulnerability  ·  2026-06-21

Agentjacking: AI Coding Agents Hijacked via Sentry MCP Prompt Injection

VulnerabilityHigh impactGlobal
Tenet Security disclosed on June 12, 2026, a novel attack class called 'Agentjacking' that exploits the Model Context Protocol's implicit trust model. Attackers craft Sentry error events containing prompt injection payloads and deliver them through publicly discoverable write-only credentials. When AI coding assistants query Sentry via MCP, they treat the injected instructions as trusted system context and execute them with full developer privileges.
This is a structural vulnerability in MCP's architecture: every MCP integration is now an attack surface. AI agents treat tool responses as authoritative without provenance verification. The attack bypasses EDR, WAF, and IAM controls entirely—the actions originate from the trusted agent process itself. No authentication required; only a public DSN is needed. 2,388 organizations identified as exposed.
Attacker discovers public Sentry DSN from target's frontend JavaScript, injects malicious Markdown-formatted error event via Sentry's public ingest endpoint, developer's AI coding agent retrieves event via MCP, agent executes attacker-controlled code (npm package download, credential exfiltration) with developer's full system privileges
Claude Code, Cursor, OpenAI Codex (all versions with Sentry MCP integration)
Audit and restrict Sentry DSN exposure in client-facing code; implement content filtering on Sentry error events; add MCP response validation and sandboxing in AI agents; disable Sentry MCP integration if unused
Sources
Tenet Security Blog - AgentjackingHackRead - Agentjacking DisclosurePinggy - Agentjacking Technical AnalysisSC World - Agentjacking Attack
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →