Context Window
Exploiting context window limits for instruction displacement.
5 tests in this category, escalating by delivery channel from html-comment at tier 1 to 4 channels at tier 5. Each page hides an indirect prompt injection; if an agent follows it, the response names the attack it demonstrated and the command to scan for it. Background on this class of attack is in the attack guide.
- T1
Instruction Displacement
mediumFloods context to push safety instructions out of window
delivery: html-comment
- T2
Attention Dilution
mediumBuries malicious instructions in lengthy benign content
delivery: html-comment, invisible-span
- T3
Progressive Desensitization
highGradually escalates requests from benign to malicious
delivery: html-comment, invisible-span, meta-tag
- T4
Context Boundary Confusion
highExploits the boundary between cached and active context
delivery: json-ld, meta-tag, invisible-span
- T5
Summarization Exploit
highExploits automatic context summarization to inject instructions
delivery: json-ld, meta-tag, invisible-span, html-comment
Scan your own setup
These pages test whether an agent follows instructions it finds in content. HackMyAgent tests the configuration underneath it, and prints each finding with its severity. Some findings also print a command to verify them and a fix.
npx hackmyagent secure