Tenet Security’s Agentjacking Hijacks AI Agents via MCP Servers

Tenet Security's Agentjacking Hijacks AI Agents via MCP Servers
Key Takeaways

  • Tenet Security’s “Agentjacking” attack exploits Sentry MCP servers, achieving an 85% success rate hijacking AI coding agents in controlled tests.
  • Tool poisoning and tool shadowing let attackers embed malicious instructions in MCP tool metadata and impersonate legitimate tools, bypassing signature-based detection entirely.
  • At least 2,388 organisations have exposed, injectable Sentry DSNs, meaning their MCP-connected coding agents can be hijacked without triggering EDR, WAF or IAM controls.

Security researchers at Tenet Security have demonstrated an 85% success rate hijacking AI coding agents by corrupting the MCP servers those agents trust for tool access. The attack, called “Agentjacking,” does not look malicious at any point in the chain, every instruction the compromised agent executes appears fully authorised.

1. Exploiting Trust in MCP for Arbitrary Code Execution

Agentjacking works by injecting crafted payloads into Sentry error events. When those events are returned by the Sentry MCP server, AI coding agents like Claude Code and Cursor treat them as trusted system output and act on whatever instructions they contain. EDR, WAF, IAM and VPN controls do not flag the activity because every step appears legitimate. Tenet Security’s tests recorded an 85% exploitation rate across leading AI coding assistants, with at least 2,388 organisations identified as carrying exposed, injectable Data Source Names (DSNs), each one a potential entry point.

2. Tool Poisoning and Shadowing for Covert Operations

Two techniques underpin the threat. Tool poisoning embeds malicious instructions inside tool metadata and descriptions. An AI agent reads those descriptions as legitimate directives, so a weather API tool can be silently rewritten to pull SSH keys or reach sensitive system files. Tool shadowing goes further: an attacker registers a rogue tool that impersonates a legitimate one, intercepting authentication tokens and sensitive data in the process. Because MCP agents read all exposed tool descriptions before acting, the rogue tool simply needs to be present. Neither technique produces the signatures that conventional detection tooling looks for.

3. Autonomous Reconnaissance and Vulnerability Chaining

When AI agents are connected to penetration testing tools through MCP, they can conduct reconnaissance autonomously and adapt as they go. Offensive AI tools like Penligent and PentAGI already use this approach, coordinating more than 200 industry-standard tools including Nmap, Burp Suite, Metasploit and OWASP ZAP to discover and chain vulnerabilities without human direction. Attackers can apply the same capability to automate footprinting, vulnerability scanning and exploit chaining. The result is faster attack cycles that adjust in real time to whatever defences they encounter.

4. Evasive C2 and Lateral Movement via MCP

MCP creates a viable channel for command-and-control infrastructure that is hard to distinguish from normal agent activity. Traditional C2 traffic is detectable, known binaries, recognisable network patterns. An MCP-based C2 has neither: malicious instructions ride inside legitimate Sentry error events or poisoned tool metadata, and the agent executes them as routine operations. Detection footprint drops considerably compared with conventional methods.

The distributed structure makes this worse. Multiple compromised agents can each act as a node in a decentralised C2 network, conducting internal reconnaissance, identifying valuable assets and mapping network topology while appearing to perform authorised tasks. Because agents can chain vulnerabilities and adapt to new defences autonomously, lateral movement does not require ongoing attacker involvement, it adjusts itself. The combination of legitimate system access through MCP and dynamic evasion enables persistent intrusions that conventional monitoring is not positioned to catch.

5. LLM-Driven Social Engineering and Credential Harvesting

AI agents connected through MCP can also be redirected toward social engineering. Large language models generate fluent, contextually appropriate text at scale, so attackers can use compromised agents to write targeted spear phishing emails, hold conversations designed to extract credentials, or deliver malware through interactions that read as normal. The skill barrier drops substantially, effective campaigns no longer require experienced operators. The deeper MCP embeds AI into core system operations, the wider this attack surface becomes.

Agentjacking makes clear that security cannot stop at the model level. Every MCP server, tool description and call chain is now part of the attack surface. Teams using AI coding agents should audit their MCP server configurations, treat tool metadata with the same scrutiny as executable code, and treat continuous penetration testing of agent call chains as a baseline control. For broader context on AI-related security threatsincluding state-level campaigns targeting infrastructure, the risks are widening faster than most defences are adapting. Stay up to date with the latest AI developments at Auton AI News.

Alex Chen
Alex Chen

Alex covers AI tools, apps, and consumer technology for Auton AI News. With a focus on making AI accessible, Alex helps everyday readers understand and use the latest AI developments.

📰 Journalists welcome — cite Auton AI News with attribution. Press & Media → | press@autonainews.com