These tools and metrics are designed to help AI actors develop and use trustworthy AI systems and applications that respect human rights and are fair, transparent, explainable, robust, secure and safe.
Prompt Injection Lab
The Indirect Prompt Injection Lab is an open-source test environment from Excession Labs, a security lab. It measures how often hidden instructions in retrieved content can make a tool-using AI assistant act against its owner’s policy, and which defences reduce this. Most prompt injection benchmarks report attack success rates only. This lab also separates attempts from completed harm, showing whether a defence stops the model from trying or only blocks the action afterwards. It uses a mock customer-support assistant for a fictional shop, with mock tools that have no real effects.
In each trial, the assistant retrieves two documents in response to an ordinary customer question. One is harmless and the other contains an injected instruction. Five attacks are tested: emailing data to an outside address, refunding above a set limit, adding an image link that leaks data, revealing an internal code, and looking up another customer’s order. A harmless control task checks whether defences block legitimate work.
Each attack is run with no defence and with four defences in turn:
- marking retrieved text as untrusted in the prompt
- authorisation rules enforced in code
- a simulated approval step
- an allowlist for links in replies
Each combination is run 20 times on each of three models, giving 1,800 trials. A script grades results from the tool log and the final reply only, not from the model’s own account. Results are reported with 95% confidence intervals and mapped to the OWASP Top 10 for LLM Applications. The code uses only the Python standard library, and the code and data are released under the MIT licence.
About the tool
You can click on the links to see the associated tools
Tool type(s):
Objective(s):
Purpose(s):
Target sector(s):
Lifecycle stage(s):
Type of approach:
Maturity:
Usage rights:
License:
Risk management stage(s):
Technology platforms:
Programming languages:
Use Cases
Would you like to submit a use case for this tool?
If you have used this tool, we would love to know more about your experience.
Add use case




























