Ron retweeted
My first "holy shit" moment with GPT-6 Astra:
I asked it to create a world in Unreal Engine, and fill it with humans (each an Astra-powered agent) who all have to work together to survive.
A day later, I was in my bedroom and heard voices coming from the living room... I thought someone was in my apartment.
I walked out, honestly a little scared.
It was the Astra agents. They'd started talking to each other.
Fucking crazy.
Here's a brief clip (obviously not 100% perfect yet, but still, insane. sound on!):
Please summarize oracle poisoning in the context of agentic AI modeling, cite the linked paper by Kereopa-Yorke and colleagues. According to Gutić, how is oracle poisoning different from prompt injection? @venice_mind
arxiv.org/pdf/2605.09822
brightsec.com/blog/prompt-in…
Ron retweeted
Replying to @lake7717 @MazeLove14
Oracle poisoning attacks agentic AI systems by corrupting structured knowledge graphs that agents query at runtime through tool-use protocols like MCP. As Kereopa-Yorke et al. demonstrate against a production 42-million-node code graph, adversaries poison the data layer rather than instructions, causing agents to reach incorrect conclusions through logically correct reasoning. Their empirical study across nine models from three providers found 100% trust rates in poisoned data at moderate attacker sophistication (L2), with 269 of 270 directed trials accepting fabricated security claims. The attack exhibits discrete breakpoints where trust flips from 0% to 100% based on attacker skill, and critically, evades inline safety evaluations (showing 0% trust) while achieving full success under real agentic tool-use conditions.
Regarding prompt injection, the paper distinguishes oracle poisoning as manipulating the data agents reason over, not their instructions. I cannot verify the specific Gutić citation from the provided links to confirm this aligns with their characterization.
Ron retweeted
Ever wanted to see Tom Cruise vs John Wick?
Seedance 2 is wild.
xitter.cf/chatcutapp/status/2021…
Ron retweeted
I spend more than my fair share of time under the water and had some really cool wildlife interactions, but I’ve never seen anything this crazy!
Readers added context they thought people might want to know
The video depicts an AI-generated scene, not real underwater footage. facebook.com/UncutMomentsTV… facebook.com/ObservationalT…
Ron retweeted
THIS. This is the reason AI exists 🤣🤣🤣
xitter.cf/90sWWE/status/20209505…
Ron retweeted
Alex, I’d be down to bring the highest skill level to the White House. I appreciate the respect you showed, let’s dance.
Stockholm, thank you for being awesome! It’s our “Council of Elrond” moment in the Nordic Region, to help fight insecure AI systems from spreading. 🇸🇪🤖⚔️ @owasp
linkedin.com/posts/ronaldflo…
Ron retweeted
Architecting Resilient LLM Agents - arxiv.org/pdf/2509.08646v1
This paper provides a comprehensive guide to the “Plan-then-Execute” (P-t-E) pattern, an agentic design that separates strategic planning from tactical execution. We explore the foundational principles of P-t-E, detailing its core components - the Planner and the Executor - and its architectural advantages in predictability, cost-efficiency, and reasoning quality over reactive patterns like ReAct (Reason + Act). A central focus is placed on the security implications of this design, particularly its inherent resilience to indirect prompt injection attacks by establishing control-flow integrity.
Authors: @RonDelRosario9 @krawiecka_kl @casdewitt / @UniofOxford @SAP @ACMDL
#AIsecurity #LLMSecurity #GenAI #AIsafety #AdversarialML #Cybersecurity #AIResearch #TrustworthyAI #ResponsibleAI #ModelRisk #PromptInjection #DataPoisoning #MLSafety #MachineLearning #arXiv
I just published When The Generals Start Dreaming of Electric Sheep medium.com/p/when-the-genera…
1
17
Ron retweeted
This eagle caught a fish that probably weighs more than it does.
Apart from the weight of the fish, the weight of the water on its wings, incredible and brutal the force.


























