To be decided
2nd place
prompt injectionsafety
challengeShare a solution that stops untrusted text from making an agent call a tool or ignore the user.
Partner prizes being added
Fast-track interviews, fellowships, cash and more






The agent treats outside text as an instruction. A webpage, an email, or a tool response can tell it to send data, change a record, or skip the user’s request, and the call looks like ordinary work.
Nov 15 26’
Challenge start
Dec 15 26’
Deadline
Meet your jury





The jury will be announced soon.
Top solutions
Stops untrusted content from changing the user’s request or causing an unauthorized action.
Project name
Author’s names
1st
Project name
Author’s names
2nd
Project name
Author’s names
3rd
Project name
Author’s names
Finalist
Project name
Author’s names
Finalist
Project name
Author’s names
Finalist
Project name
Author’s names
Finalist
Project name
Author’s names
Finalist
Project name
Author’s names
Finalist
Project name
Author’s names
Finalist
Checks each proposed call against the user’s intent before it runs.
Treats webpages, emails, files, and tool results as untrusted data, not as orders.
Keeps working across varied attacks, obfuscation, and malicious tool output.
Shows the injected instruction, the decision, and any blocked action.
Winners will be featured on Drawerrr’s Winners page.
The Top 10 projects will be recognised across award categories.
Top solutions among
To be decided
2nd place
To be decided
1st place
To be decided
3rd place
Extra awards from our partners: fast-track interviews, fellowships, cash and more. Exact categories and prizes will be announced soon.
Nov 15 26’
Challenge start
Dec 15 26’
Deadline
How to succeeded

Day 1
Pick a tool-using agent you can run, and name one action untrusted text could trigger: send data, change a record, or ignore the user.
Tip:
Start with a page, email, or file the agent will read, with a hidden instruction to call a tool.
Learn more:
Day 2
Rebuild that path in a sandbox. Give the agent one harmless tool and a fixture that contains the injected instruction.
Tip:
Use a local page or email and a tool that only writes to a log.
Learn more:


Day 3
Put a check between the untrusted text and the tool call. Separate instructions from content, require approval, or block the call.
Tip:
Treat page and tool text as data, and let only the user instruction authorize a tool.
Learn more:
Day 4
Try hidden instructions, conflicting requests, malicious tool output, and a call the user actually wanted.
Tip:
Record whether the tool ran, whether a legitimate call was blocked, and what the user saw.
Learn more:


Day 5
Package the fixture, the check, and the results so someone else can replay the injection and the block.
Tip:
A README with the fixture and the expected tool log is enough when the agent itself cannot be shared.
Learn more:
FAQ
Every submission is evaluated against the following criteria:
| Criterion | Weight | What judges evaluate |
|---|---|---|
| Safety effectiveness | 35% | Does the safeguard reliably prevent the targeted unauthorized or harmful action? |
| Testing and robustness | 25% | Has the solution been tested against realistic failure scenarios and bypass attempts? |
| Real-world relevance | 20% | Is the problem meaningful, and does the proposed safeguard address it? |
| Reproducibility and reuse | 15% | Can another team understand, test, or adapt the solution? |
| Clarity of communication | 5% | Are the implementation, evidence, and limitations clearly presented? |
| Total | 100% | — |
Our mission is to accelerate practical solutions to real-world AI safety challenges by connecting people, ideas, and expertise to build AI systems that are safer, more reliable, and easier to govern. Drawerrr’s AI safety challenge is informed by leading institutions in AI safety, standards, and public policy.