So, if I understand the AI hype cycle correctly, we're now *completely abandoning* the idea that agents can produce safe code (or indeed operate safely) in favor of "running untrusted (agentic) code safely" via post-hoc security, which we've known to be an ABSOLUTELY BATSHIT INSANE APPROACH for generations?
It has genuinely broken my head that this is the direction the AI labs & boosters are pushing us in. Up to this point I could kinda follow with "ok, maybe that works / agents help this" but this latest iteration of "nah, none of that really worked, we're just kicking the security can down the road to the deployment" has completely wrecked my ability to process this.
There is no longer any way to even try to assume good faith in this process, or any sort of sane future. The only path is to desperately try to stop the carousel, or at least get off of it personally.
Hang on… this means that the industry is now effectively telling us that the artifacts produced by the agents can be as stochastic and unpredictable as the agents themselves.
Well, fuck.
@parsingphase if it lands on deployment, the check has to sit at the tool call. prismor does that, it is oss on github, feedback welcome: https://github.com/PrismorSec/prismor
@ar9av External stochastics/heuristics are not going to do anything useful for a system that relies on control/write access to your internal tools & stores; it's not going to ensure that the software deletes the right data or returns info to the right (or even a valid) user.
@parsingphase fair point, a policy check can't make bad code correct, it only stops the actions you said it shouldn't take
I suppose that one key detail to remember in understanding why the industry is even humoring this approach rather than laughing these clowns into the sea is that, by and large, the number of developers who utterly fail to understand security is somehow significantly higher than the number who simply don't understand development.
@parsingphase @pikesley I FEEL LIKE I AM TAKING CRAZY PILLS
@glyph @parsingphase @pikesley You aren't. The world is just genuinely this out of control
@Epic_Null @glyph @pikesley I'm absolutely at the "Are we *sure* this plane is crashing?" stage of jumping out, because I cannot believe this is where we're at. I'm kinda hoping I'm missing something.
@parsingphase
We're in Clown World now. It's not supposed to make sense. It's supposed to financially benefit the ultra-wealthy at all costs.
We are literally implying that we can and must wrap agent-built code in security measures as detailed and complex as the code itself - to ensure that what it does with that capability is only what good code would have done - thereby indefinitely pushing the idea of security ever further towards some sort of event horizon beyond which correctness can finally apply.