Darshana

Senior Eng @ FAANG · AI infra. How AI agents can be secured, and kept under control

[Darshana @Darshanaxyz]

@0xMorlex Great write up! It's super important to move the evaluation and guardrails out of the LLM/agent.

[Darshana @Darshanaxyz]

@Remybuilds How so?

[Darshana @Darshanaxyz]

@anchorstack_dev Langchain for observability for now, guardrails on my own. How about you?

[Darshana @Darshanaxyz]

@IAmSandroSaric Done!

[Darshana @Darshanaxyz]

@souvikdeb26 sg, telemetry is super important!

[Darshana @Darshanaxyz]

@sobedominik I could be wrong but i feel like this has nothing to do with AI

[Darshana @Darshanaxyz]

Instruction based guardrails for agents fail for a simple reason: the constraint and the attacker live in the same context window. A prompt that says "don't exploit, only report" is text the model can be argued out of. Safety that is enforced only inside the prompt is not really a guardrail.

[Darshana @Darshanaxyz]

@damonchen Curious to hear, do they not all look the same?

[Darshana @Darshanaxyz]

@BoasCFC @gregisenberg https://t.co/ihNQAK3DJb is a good one and https://t.co/8hASIfM4kS is also a good one

[Branko @brankopetric00]

this is fine for vibe-coded projects with non-critical production workloads, where downtime doesn't actually affect your users or you.

nobody's losing sleep if the site is down for 10 seconds.

but this can't be applied to projects with an SLA, where every minute you're down literally costs you money.

we've already seen examples of ai agents breaking critical features in products used by the whole world.

and once that happens it's not just the downtime, it hits your reputation and it changes what users expect from you.

there's a reason big companies run staging environments and don't let anything touch prod directly.

it's not because they're slow or scared, it's because the cost of getting it wrong is real.

in general this whole approach is fine for tiny little saas projects, mainly b2c, where you're the only one who feels the pain if something breaks.

anything bigger and you need guardrails.

[Darshana @Darshanaxyz]

@pcshipp telling ai what to build

[Darshana @Darshanaxyz]

@icanvardar totally second this. everything about agentic workflows from prompt best practices to guardrails to deployment is still evolving.

[Darshana @Darshanaxyz]

@TheSpacerr totally agree

[Darshana @Darshanaxyz]

@de1lymoon Guardrails is super important step that lot of folks skip.

[Darshana @Darshanaxyz]

@gabriel1 looks like moist is universal, this is claude's response

[Darshana @Darshanaxyz]

Takeaway: treat output correctness as a first-class signal, not a postmortem artifact. Until you can page on "the agent got dumber," cost and latency graphs are measuring the wrong thing precisely.

[Darshana @Darshanaxyz]

Production AI agents have decent tooling for cost and latency but not much for quality. You can prove your agent is fast and cheap, you can't prove it's still correct. This is where silent degradation happens.