Antoine Carossio

At that time I was a teenager fascinated by @Apple, dreaming of revolutionizing the world. 10 years later my passion and determination are always stronger. You inspire me in each of my decisions. Thank you Mr Steve Jobs. #RememberingSteve

Your voice shouldn't need a cloud subscription. 🎙 Voicebox runs voice cloning, speech generation and Whisper dictation locally. Free, MIT-licensed. Its MCP server lets Claude Code, Cursor or Cline use voice as a tool. 🔥 Your mic shouldn't become someone else's API dependency.

Give GPT-6 Astra a lab exploit task and it refuses. Give GPT-6 Cyber the same and it completes 95%. ⚡ @OpenAI RemoteExploitBench: Astra/Daybreak Blue refused 100%; Cyber hit 95%. 🔥 Zero-days used to be bottlenecked by human hours. Now it's compute and who makes the allowlist.

Nobody: what's the most SF thing you can do? Me:

@deepseek_ai just open-sourced dsh-superagent Always-on coding agent: watches repo + team chat, picks work. Same as my linear-autopilot (7d, 123 tickets). Fix first: full OS access by default. Limited not by intelligence, but by who gets to talk to them. github.com/deepseek-ai/dsh-s…

A 284B-parameter model on a single gaming GPU. 🖥️ FreeToken, from a Berkeley team incl. Ion Stoica and @matei_zaharia, parks MoE experts in RAM and pulls only what each token needs over PCIe. Catch: 192GB+ of RAM, and the API key is off by default. github.com/FlashML-org/FreeT…

A public benchmark has a short shelf life: it becomes training data. To test Cascade 2 Iguazu, we built our own at @escapetechhq. Critical findings: 18 → 45, mostly from coverage. Looking everywhere beat thinking harder. Breakdown by Arnaud Fanthomme: escape.tech/blog/ai-pentesti…

Google finally takes the lead, and the biggest gap is in cyber. 🛡️ Gemini 4 Carbon vs GPT-6 Astra: - Finding vulns: 88.2% vs 85.4% - Exploiting: 54.8% vs 42.4% - Patching: 71.0% vs 68.0% New frontier model every 2 weeks, for attackers too. Is your pentest still once a year?

Want to learn security in 2026? Stop collecting certificates. Break your own stuff instead. 🔓 build: a login flow, an API, an agent attack: replay a token, change an ID, trick the prompt fix, attack again detect: check the logs Tools change. This loop doesn't. @escapetechhq

Today we're releasing Cascade Iguazu at @escapetechhq. We rebuilt our AI harness from the ground up: systematic coverage, full auditability, better vulnerability discovery. AI pentesting security teams can actually trust. Breakdown and benchmarks: escape.tech/blog/introducing…

Twice the screen. Twice the attack surface? 😜 #iPhoneDuo @escapetechHQ #NewYork #5thAvenue

Every PHP version since 5.4 can be hacked with zero login. 😅 php2shell: 2 critical CVEs in the PHP engine itself. Not WordPress, not your code. PHP. No fix for 8.1 and older, and exploits seen before the patch shipped!! Live in @escapetechHQ ASM: escape.tech/blog/php2shell-c…

🚨 HUGE MILESTONE: Cascade can now hack your AI agents. 😅 Today we launched AI Red Teaming in @escapetechHQ. Agent, MCP server, APIs: each tested with its own tool. Nobody tests the path between them. That's where critical vulns slip through! Live for all AI Pentesting users.

Attackers have AI agents running 24/7. Most companies still pentest once a year. 😅 @opencode just got breached. Mid-incident, the models they used to investigate started refusing. Defenders hit the guardrails, attackers didn't care. Would your stack survive 48h of that? 👇

People are now vibe coding their TV 📺 @gathint had Claude Code strip the ads off his parents' TV. 48 hours later: 5,118 GitHub stars, 618 forks and 54 TVs tested. The security guy in me: you're giving an AI agent debug access to your home network. Read what it runs!!

Guess what is the relation between VR and Security? 👨‍🍳 @escapetechhq

In NYC for the next 2 weeks, let's grab a coffee? ☕️