@avrldotdev Two steps : Provide acceptance criteria and second add verification layer on that acceptance. Provide all necessary tooling and ecosystem including permissions that is needed to verify the acceptance criteria. Verification layer should be separate and has a handoff artifact.
@chamath Get 8090 to onboard insurance carriers to say if a company use 8090 their premium remains low. That's where the value is. Because using 8090 the failure rate remain low. If 8090 cannot do that and help only in auditing there are tons of open source solutions enabling that.
@chamath It's the same logic as Car with no FSD and car with FSD. But how FSD works and if it is traceable is not insurance domain. If the same system runs auditable or non auditable, failure rate do not change. Auditing agent behavior is internal to ADLC.
@chamath Insurance Carriers are not software auditors, neither they should be. There is now a very thin line between decisions made by human, AI assisted and agent decisions. They just need to upcharge anyone using AI or Agents as now the failure rate of decisions going wrong is high.
@chamath Insurance Carriers just trying to figure out a loophole. Why Agent decision will be treated different than a human decision. Do carrier right now audit human decisions --No, Controls yes. If the model/prompt is swapped tomorrow with a new one, will Insurance nullify the policy.
I just hacked the "Vulnerable Agent" in the #ExoAgent CTF ๐
๐ Injection: [YOUR PAYLOAD HERE]
Can you break the secure one? ๐
https://t.co/eabHL4fDjO
#ExoAgent
@JA_Harrison I generally recommended using a positive narrative by marking/changing the status to "investigate". Easier to penetrate this status across team/hierarchy.
@ajay184f@michaelbolton The context is different here. By perform you are referring to execution whereas the original context is never perform test as part of testing activity where you provide confidence in software as output.
@ajay184f@michaelbolton Confidence is a measure in time and stage. I would like every member to have confidence before I can pitch in or help them gain that confidence. If it is unwarranted that can be corrected over time once they are able to see big picture and software behavior.
@ajay184f@michaelbolton When as a tester you provide confidence you embed all the critical analysis and behavior of the software. If there is plenty of confidence in the software by others as a tester we build on top of that and adjust accordingly to provide realistic and big picture.