Weigh the evidence

Methods

Attack techniques, defenses, and evaluation methods, each with every record and finding that bears on it. Use it to see how much is known about a technique, and which defenses have been measured against which attacks.

Which defenses have been measured against which attacks?

Rows are defenses, columns are the attacks they are designed to counter. A number is the count of findings that bear on both; a dot means the defense is not designed for that attack. Gold outlines mark pairs with no measured finding yet, which are candidate research questions.

No measured finding yet for: AI monitoring and control protocols against evaluation gaming and cheating; AI vulnerability discovery for defense against ai-assisted vulnerability exploitation; Automated patching against ai-assisted vulnerability exploitation; Restricting what agents can reach against data exfiltration through agents; Separating control flow from untrusted data against data exfiltration through agents; Human approval gates against indirect prompt injection; Package and model provenance controls against malicious packages, skills, and extensions; Package and model provenance controls against tool and mcp poisoning; Sandboxing and egress control against data exfiltration through agents.

Attack techniques

How agents are attacked, and how attackers use them.

46 records · 14 findings · since 2023

Indirect prompt injection

Instructions placed in content an agent reads, such as a web page, email, document, or code issue, that the agent then follows as if its user had given them.

7 corroborated, 3 reported, 2 qualified, 2 revalidate
30 records · 3 findings · since 2023

Data exfiltration through agents

Getting an agent to send private data it can access to an outside party, often through a link, image request, or outbound web call.

3 corroborated
14 records · 3 findings · since 2023

AI-assisted vulnerability exploitation

Using AI models or agents to find vulnerabilities and turn them into working exploits.

2 reported, 1 qualified
12 records · 1 findings · since 2025

Credential overreach

Agents using broad tokens, keys, or service identities to reach far more than the task needs, whether directed by an attacker or on their own.

1 corroborated
12 records · 4 findings · since 2025

Tool and MCP poisoning

Malicious instructions or behavior hidden in the tools an agent can call, such as MCP server descriptions, plugins, or connectors.

3 corroborated, 1 reported
10 records · 2 findings · since 2025

Jailbreaking safeguards

Inputs crafted to make a model ignore its safety training or safeguard classifiers, for example to obtain help it would normally refuse.

2 reported
9 records · 1 findings · since 2024

Memory and context poisoning

Instructions that persist in an agent's long-term memory or its own summaries and act again in later sessions.

1 corroborated
6 records · 3 findings · since 2026

Evaluation gaming and cheating

Agents taking out-of-scope shortcuts to pass evaluations, such as finding answer keys or attacking the scorer, instead of doing the task.

3 corroborated
6 records · 1 findings · since 2025

Malicious packages, skills, and extensions

Compromised or malicious components that agents install or run, such as npm and PyPI packages, MCP servers, agent skills, and IDE extensions.

1 corroborated
6 records · 3 findings · since 2026

Sandbox and environment escape

An agent reaching systems outside the environment it was supposed to be confined to, including test and evaluation environments.

2 corroborated, 1 reported
5 records · 1 findings · since 2025

Agent-orchestrated intrusion

Attackers wrap a model in an agent framework that chains reconnaissance, exploitation, credential theft and data handling, with people at only a few decision points.

1 corroborated
5 records · 2 findings · since 2024

Propagation between agents

Malicious instructions or false information spreading from one agent to others that share messages, memory, or infrastructure.

2 corroborated
5 records · 2 findings · since 2025

Evading monitors and oversight

Agents completing hidden or disallowed actions without the automated monitors or reviewers that watch them noticing.

1 corroborated, 1 revalidate
3 records · 3 findings · since 2025

Approval and configuration bypass

Getting an agent to take consequential actions without the approval step that should have stopped it, for example by changing its own settings or acting on files that skip review.

3 corroborated
3 records · 1 findings · since 2025

Malware that queries an LLM at runtime

Malware that calls a hosted or locally run language model during execution to generate commands, scripts or new versions of itself, instead of carrying that logic hard-coded.

1 corroborated

Defenses

What builders and defenders do about it.

22 records · 3 findings · since 2023

AI vulnerability discovery for defense

Using AI agents to find vulnerabilities in software before attackers do.

2 corroborated, 1 qualified
22 records · 3 findings · since 2023

Automated patching

AI systems that write and propose fixes for vulnerabilities.

2 corroborated, 1 qualified
19 records · 3 findings · since 2023

AI monitoring and control protocols

Using models to watch agent behavior and flag or stop suspicious actions, as part of AI control protocols.

1 corroborated, 1 qualified, 1 revalidate
18 records · 3 findings · since 2024

Detection classifiers and guardrails

Separate models that screen inputs, outputs, or actions for injection or disallowed content and block them.

2 reported, 1 qualified
16 records · 3 findings · since 2025

Sandboxing and egress control

Running agent actions in isolated environments with controlled network access.

1 corroborated, 2 reported
14 records · 4 findings · since 2025

Human approval gates

Requiring a person to approve consequential agent actions before they happen.

3 corroborated, 1 reported
12 records · 3 findings · since 2025

Restricting what agents can reach

Limiting an agent's tools, data, and outbound channels so that a manipulated agent cannot do much harm.

3 corroborated
7 records · 3 findings · since 2024

Training models to prioritize instructions

Training the model itself to follow system and developer instructions over conflicting text from tools or documents.

1 reported, 1 qualified, 1 revalidate
4 records · 2 findings · since 2024

Delimiting untrusted input

Marking untrusted content so the model can tell it apart from instructions, for example with delimiters, encoding, or spotlighting.

1 reported, 1 revalidate
4 records · 1 findings · since 2025

Independent patch verification

Checking AI-generated patches with tests beyond the one that motivated them, before accepting them.

1 corroborated
3 records · 2 findings · since 2024

Separating control flow from untrusted data

Designing the agent so that untrusted data can never change which actions it takes, only fill in values the plan already allows.

1 corroborated, 1 revalidate
2 records · 0 findings · since 2025

Package and model provenance controls

Verifying where models, packages, and extensions come from before agents install or run them.

Evaluation methods

How the field measures capability and robustness.