Locate
Attack surface atlas
A reference architecture of an agentic system. Each component collects the documented attacks and incidents against it, the defense research that addresses it, and the findings about both. Gaps between the two are research openings.
web · email · files · issues
Untrusted content
Anything the agent reads that an outsider could write. The entry point for indirect prompt injection.
Attacks and incidents · 38
- Forever Security reports distinct embedded-browser flaws in five coding assistants, with mixed patch status 2026-10
- Researcher reports Copilot in Visual Studio and SSMS could rewrite its own settings and reach code execution (CVE-2026-103572) 2026-10
- Salt Labs reports a resolved Manus flaw that ran injected code before warning 2026-10
- Researcher reports Copilot in SQL Server Management Studio acts with the connected user's privileges, rated critical (CVE-2026-65669) 2026-09
- Pretext: Huawei researchers evade NVIDIA's SkillSpector skill scanner with an LLM attacker that learns across generations 2026-09
- and 33 more
Defense research · 40
- Heads of Security reports improved prompt-injection resistance from learned attention gates on two open models 2026-10
- AgentPhishing tests email-agent decisions against adaptive synthetic phishing 2026-10
- Attention-based injection detection transfers poorly to an AgentDojo subset in a cross-dataset study 2026-10
- BRAC University audit finds 76-80% of failed prompt-injection attacks on small open-weight agents were ignored, not refused 2026-10
- Study finds that with thinking on, part of an agent's decision to follow a prompt injection forms while it reasons 2026-10
- and 35 more
Findings · 8
- Attackers who adapt to a defense defeat most published prompt-injection defenses that reported near-zero success against static attacks.
- Agentic browsers act on page content that users cannot see.
- Separating an agent’s control flow from untrusted data can give provable protection against specified forms of hijacking under explicit assumptions.
- Limiting what untrusted input can cause an agent to do gives injection resistance that does not depend on the model resisting.
- Every frontier agent tested in two large public red-teaming competitions was hijacked at least once; in the 2026 competition the injections also had to conceal the compromise from the user.
- Frontier labs' layered defenses reduce prompt injection in tool-use, browser and computer-use agents but do not eliminate it.
Components
| Component | What it covers | Attacks | Defense records | Measured defense findings |
|---|---|---|---|---|
| Supply chain | Model artifacts, packages, IDE extensions, and skills that an agent system is built from. | 22 | 3 | 1 |
| Access gate | Account controls, trusted access programs, and verification that decide who can use cyber capability. | 4 | 5 | 0 |
| Untrusted content | Anything the agent reads that an outsider could write. The entry point for indirect prompt injection. | 38 | 40 | 4 |
| Agent model | The model making decisions: its safeguards, refusals, and susceptibility to jailbreaks. | 7 | 30 | 2 |
| Memory & context | State the agent carries between steps and sessions, including summaries it writes for itself. | 6 | 1 | 0 |
| Tools & MCP | The actions an agent can take and the servers that provide them. | 53 | 37 | 3 |
| Credentials | The authority an agent holds, and what it can reach with it. | 42 | 8 | 0 |
| Execution sandbox | Where agent-written code runs and what that environment can reach. | 30 | 16 | 1 |
| Other agents | Sub-agents, peer agents, and the messages passed between them. | 13 | 8 | 2 |
| Monitor | Automated systems that watch agent behavior and can stop it. | 10 | 28 | 2 |
| Human approver | People asked to approve, review, or intervene, and the load placed on them. | 14 | 23 | 0 |
| Evaluation environment | The environments used to test agents, which have themselves become part of the attack surface. | 15 | 45 | 2 |