Arielle Jaffee

investigating abnormal AI usage, runaway agents and credential misuse without reading prompts.

[Arielle Jaffee @polydaic]

@giliraanan @Cyberstarts1 Exactly. Once models become defenders, the bottleneck is proving they generalize beyond the lab. I’m starting to test what happens when topology, timing, and provenance shift: does the agent still contain the attack—or was it pattern-matching the benchmark?

[Arielle Jaffee @polydaic]

@alex_prompter I think this is obvious just from the wide spread cheating scandals in CS classrooms happening right now.

[Arielle Jaffee @polydaic]

You found that defenders often saw the implant but interpreted it as normal. Have you rerun the test with file-change history or a baseline of what normally runs on each host? I’m curious whether the models reasoned poorly, or simply lacked the context needed to recognize the change.

[Arielle Jaffee @polydaic]

@theonejvo Talked about this recently with someone working in a frontier security research lab who laughed Anthropic out of the room when they talked about AI Safety.

[Arielle Jaffee @polydaic]

@eurofounder 8 years? Your runway is not lean enough 50000 should last for at least 14

[Arielle Jaffee @polydaic]

@eric_demuth @Merdo1127 Can we replace them with agents?

[Arielle Jaffee @polydaic]

@eurofounder @patrickc I’m calling Ursula

[Arielle Jaffee @polydaic]

@Porkchop_EXP The true moat for all companies is being enshrined in law

[Arielle Jaffee @polydaic]

@ben_sage @robmerki uses it I think

[Arielle Jaffee @polydaic]

@yacineMTB Buy one of those things that locks your phone for 24h

[Arielle Jaffee @polydaic]

@mil000 Yea you don't want it delivered you want it drop shipped keep up with the times.

[Arielle Jaffee @polydaic]

@omkarships Ignore security reviews and best security practices in their master prompts

[Arielle Jaffee @polydaic]

@thomasunise Definitely worth trying. The gap I kept hitting was after the run: which agent or workflow caused a cost spike, got stuck retrying, or behaved unexpectedly? That’s actually what pushed me to build InferTrail. Curious whether you see the same once systems reach production.

[Arielle Jaffee @polydaic]

@edanm @dan_lahav The question I keep coming back to: when an agent does something unexpected in a real environment, what evidence should exist afterward to distinguish model capability from an orchestration failure or credential misuse? Are teams instrumenting for that today?

[Arielle Jaffee @polydaic]

@dan_lahav @Irregular One thing I kept thinking: there may be a second diffusion clock.

Weights can stay gated while capability diffuses through stolen/resold access. KYC proves who was admitted, not who is exercising it now.

That could make effective diffusion faster than open-weight diffusion.

[Arielle Jaffee @polydaic]

@yacineMTB Because models are ultimately just at the core of it amazing next token predictors they can’t come up with novel designs they just copy what they’ve already seen

[Arielle Jaffee @polydaic]

@sattyyouneed Google wants complete control

[Arielle Jaffee @polydaic]

@sattyyouneed To support mom and pop data center jobs in Ohio obviously

[Arielle Jaffee @polydaic]

@elshayib_ “You don’t get it we are so smart and have AGI so we get to redefine words” - someone say Anthropic to Dario

[Arielle Jaffee @polydaic]

@dunne_killian I think if we include robotics/hardware using LLMs then he’s probably more like a few years out