• Sources: primary, discussion
  • Summary: The announcement pairs OSS Scanner, which files model-generated vulnerability reports to open-source projects that enroll, with a Critical Infrastructure Defense Program listing 11 founding partners. Each scanner report carries a proof of concept, an explanation and a suggested fix where one is available, and no human checks it before delivery. Anthropic states an expectation of above 90 percent true positives and warns that some reports will contain inaccuracies such as a wrong severity rating.
  • Why it matters: Maintainers absorb the triage cost of unreviewed reports that the vendor itself expects to be wrong some of the time, one day after this archive published LWN's account of kernel maintainers already triaging a flood of model-written bug reports.
  • Follow-up: Check whether Anthropic publishes a true-positive rate measured on enrolled projects, rather than the stated above 90 percent expectation.