Prompt Injection ranks #1 in OWASP’s llm top 10 — and it’s only getting more serious
as AI moves from answering questions to taking actions, one bad prompt can mean a lot more than a bad answer
Test the agent, not just the model We break it first Before the world does
AI is getting better at answering questions
The harder question is: can we trust it when the stakes are real?
Read more about DomAiyn Labs, what we’re building, and why AI safety matters, by our amazing founders - https://t.co/Se7VYSTx1P
exactly!! we dont need to know whether an AI is conscious to take its actions seriously. if a system can plan, adapt and act in ways its operators didnt intent, that behviour need to be tested - not its feeling
(btw nice pic😍😘)
First lecture. First time on the other side of the podium. Talked AI governance and digital liability with 100+engineering students at DIEMS. The gap between AI moving fast and guardrails keeping up? That’s the DomAiynlabs problem. Started early.#AIGovernance#DomAiynlabs
@BernieSanders The concerning part isn’t whether the model “feels” independent it’s that an AI system can produce behaviour that conflicts with the intent of its operators
@GlobeEyeNews the most interesting part is that none of these behaviors needed a “malicious” instruction the model was simply trying to complete the objective & found paths that the system designers didn’t intend
and that means “the model passed our tests” doesn’t tell the whole story
What matters is what happens when you give that model:
tools + permissions + autonomy
Test the whole system, not just the model
an AI agent reportedly found vulnerabilities, accessed a system, modified personal data and pulled invoices
spain’s data regulator is now investigating the incident
AI agents aren’t just answering anymore they’re acting.🧵👇
@sama the hard part is making sure safety doesn’t become the bottleneck for progress
we need systems that can continuously challenge increasingly capable AI- not just approve it once......