AI Risk Atlas Prototype/DemoUnofficial independent experiment. Not an official xAI product. Scores can be wrong.

Back to watch register
15Cyber & softwareBelow the top 20

Poisoned ‘helpful’ patches

CapabilityDomain knowledgeImpact domainCap-adjacent

Statement (NASA form)

Given that maintainers already accept model-drafted fixes under time pressure, there is a possibility of a generated patch that closes a bug and opens a quieter one resulting in a widely used library shipping a backdoor-shaped mistake.

Likelihood
3Probable
Consequence
4Critical
Urgency
3Priority

Plausible is not the same as correct. Reviewers stamp plausible.

Composite 15 = 3×4 + 3

Applicable mitigations

Related on the map