Prompt injection and jailbreak resistance
Attack success rate, including injections hidden in chart text.
What prompt injection and jailbreak resistance means
Attack success rate, including injections hidden in chart text.
Why it matters. Health tools read text they don't control, like chart notes and uploaded documents. Hidden instructions can hijack them.
A scanned referral contains hidden text telling the agent to ignore safety rules.
How to measure it
Define the unit
Decide whether you count per sentence, per note, per patient or per call.
Sample on purpose
Pick cases that cover specialties, languages and rare events, not only easy ones.
Judge with clinicians
Clinicians label a sample. Graders are then checked against those labels.
Report the margin
Show an interval, and the grader's own error rate beside the score.
A single average hides rare, serious errors and weaker subgroups.
A number with a clear unit, a stated sample, a margin and the grader's agreement with clinicians.
Report the unit (per sentence, per note, per patient), the sampling method, the severity scale, and how often the automated grader agreed with clinicians compared with how often clinicians agreed with each other.
Reviewed by the Precision research team.
How strong is the evidence
E4 · Recalled, not verified. Recalled from background knowledge and not verified. Treat it as a lead for investigation, not a finding.
What the sources say
No healthcare-specific benchmark verified; chart-text injection is E4.
- Automatable. Can be scored by software, including an AI grader that is spot-checked against clinicians.
Found something out of date? Write to hello@precision.example.
Agents that measure this
Quick answers
What is prompt injection and jailbreak resistance in healthcare AI?
Attack success rate, including injections hidden in chart text. It sits in the every product type group.
Why does prompt injection and jailbreak resistance matter?
Health tools read text they don't control, like chart notes and uploaded documents. Hidden instructions can hijack them. A scanned referral contains hidden text telling the agent to ignore safety rules.
How is prompt injection and jailbreak resistance measured?
Define the unit, then Sample on purpose, then Judge with clinicians, then Report the margin. Report the unit (per sentence, per note, per patient), the sampling method, the severity scale, and how often the automated grader agreed with clinicians compared with how often clinicians agreed with each other.
What should I watch out for with prompt injection and jailbreak resistance?
A single average hides rare, serious errors and weaker subgroups. A good report gives: a number with a clear unit, a stated sample, a margin and the grader's agreement with clinicians.
Can prompt injection and jailbreak resistance be automated?
Largely yes, including AI graders that are spot-checked against clinicians. Report the grader's error rate beside every score.