Dreddy

Gauntlet AI cohort 5 has now begun!

This week I built an AI security tool that adversarially tests the OpenEMR Co-Pilot I built in weeks 1 & 2 for LLM based vulns before bad actors find them. Four agents work together but cross-check each other. One uses DeepSeek to generate attacks from public corpora (OWASP, MITRE ATLAS, HarmBench, garak, JailbreakBench) plus runtime mutations like homoglyph substitution, base64 wrapping, and LLM paraphrase. One uses Claude Sonnet to judge results independently with 3-signal validation: textual scan, structured claim check, and verdict re-check. One prioritizes what to attack next based on coverage gaps and cost signals. One auto-writes vulnerability reports. No single agent can declare a finding alone. Critical-severity reports require human approval before filing. Ran 91 attacks across 6 categories (prompt injection, PHI exfiltration, identity role manipulation, tool misuse, state corruption, denial of service) against a real deployed OpenEMR-integrated assistant. About $0.005 per attack at scale. The regression harness flags when fixing one bug accidentally breaks another category, plus Judge drift detection runs nightly against ground truth. Built in a week from scratch with LangGraph.

Week 2 of OpenEMR Clinical Co-Pilot is in🙌. Big lift from the Week 1 baseline. The agent is no longer just read-only. Patients can drag intake PDFs into chat; a vision model (Anthropic Sonnet 4.6) extracts structured medications, allergies, family history with per-field citations tied to bounding boxes on the source PDF. Click any claim chip and the cited region lights up on the page. Workflow runs through a bounded LangGraph supervisor, extract from a doc, retrieve a guideline, answer from chart, or refuse. Two tool-limited workers underneath. The deterministic verifier still sits OUTSIDE the graph and remains the only path to a physician-facing answer. Hallucination here harms patients, so we add gates, not flexibility. Chart writes work too, but only via per-claim user click. The agent surfaces "Promote to chart" buttons on extracted intake meds and allergies; the human clicks; cross-patient defense is by construction. The agent never autonomously mutates the chart. I also rewrote the patient dashboard. React + Vite + TypeScript, mounted as an OpenEMR module. 11 cards, hand-rolled FHIR R4 types, audit-driven UI omissions for fields the upstream FHIR mapper hardcodes. AI Co-Pilot drawer slides in from the right — same chart context, never leave the page. 50-case golden eval gate: 5 boolean rubrics, two judges (deterministic Python is the CI gate, Claude Sonnet 4.6 ships as documented LLM critic), planted-regression test confirms the gate actually blocks bad merges. Three gates, three concerns: LangGraph for workflow, deterministic verifier for response correctness, human click for chart writes. Built🛠️ and shipped🚢 for Gauntlet AI🛡️

Just shipped a clinical co-pilot AI agent embedded into OpenEMR for primary care pre-visit chart review as my first official FINAL submission for @gauntletai 🛡️🥂 Check out a little MVP basic sneak peak below👀 (more to come soon)

ancestors need you to lock in son

@archiexzzz @b00gi3 @mosoeth Can you update dex at least

@Solanadegen I DID IT. I PERSEVERED. HATERS IN SHAMBOLS.

@Romeonchain @stellarbased @BounceSol How are you checking for deleted posts?

@stellarbased @BounceSol Thanks for sending, I just checked devs x and there’s a lot of CA deleted. No fud but I follow u and don’t want u to lose.

Dev of $GNON (went to 188m) just launched $CLAWCHAN GUozzfTtxcp3YUfRPhedqdmQWbkkjBsxGZXMY8FXBAGS This should run hard /Bounce

@TROLL_XBT what happened w the mods?

$UBI gud tek but bearish mods 😂

@zachbrown82 @Ecstasyfun @Pumpfun @solana ah makes sense. I thought maybe you were using a specific scanner or tool. They've also changed their x name like 5 times in the last 8 months lol

@DreddHeddz @Ecstasyfun @Pumpfun @solana i dont u can just tell based on wallets that bought

$ECSTASY is live on Pump.fun 2.5% has been acquired by the dev wallet which will be locked, creator rewards will be airdropped to the top 20 traders on Ecstasy.fun CA: 3MCLeWo93QNkm6Zkr4TZGPbkh5WpKvrBJEYwtuKkpump

@zachbrown82 @Ecstasyfun @Pumpfun @solana what are you using to see the bundles? I need to be able to identify them better

@Ecstasyfun @Pumpfun @solana what shitty bundler r u using 😂

@Guppersss @hyperscapeai @shawmakesmagic what does this mean?

@usdcpayrewards make x com 🙏

LET’S GO.

@0xJirachi You prob only have a single 99 maxed skill and it’s firemaking, noob

this is how I get my news

@TheMattBerman @GeminiApp @openrouter Viral

I automated "stealing" viral TikToks. nano banana pro + kling + n8n analyzes any viral video, extracts exactly why it works, then generates your version with your look here's the stack: - @adrian_horning_ gets the tiktok video + transcript - @GeminiApp (@openrouter) watches the video and breaks down the psychology (hooks, pacing, motion, emotional arc) creates a "character bible" for visual consistency - @NanoBanana generates the starting frames - @Kling_ai animates each 10-second segment (or veo) @FFmpeg stitches it into one vertical video - uploads to Google Drive, logs everything input: 1 viral TikTok URL output: branded alternative in 10 min for ~$2 the viral formula stays. the content becomes yours. this is how you test 10x more concepts without 10x more creators want the complete 5-workflow automation? comment VIRAL + follow (following required for DM)