Road of Becoming Founder and CIO | Learning and Studying Applied AI In University of Memphis | Emerging Creative Technologist and Content Creator | Tech and AI
Learning AI security? Try garak, NVIDIA's free open-source LLM vulnerability scanner. Think nmap, but for LLMs: it probes for prompt injection, jailbreaks, data leakage and more. One pip install, and great for students and devs testing their own apps. https://t.co/zEWlfWWmQc
Anthropic says an unreleased Claude model chained four Chrome bugs into a one-click exploit that escaped the sandbox. Google patched 62 critical or high CVEs from the work in Chrome 156. My take: AI finds bugs faster now, so update your browser today. https://t.co/nzUTnxkEiX
@empowwerr Good to connect! What kind of agents are you building right now, and are you using a framework or rolling your own loop? Curious how the distributed databases side plays into it
@asaridarr Nice, scoring posts to pick which ones to reply to is the exact problem I've been thinking about. What does it weigh most, the author's reach, how fresh the post is, or how well it fits your niche?
@TheXSami That's how I'd want it too. A pass/fail hides the reasoning and you'd still have to look it up yourself. Affected version, fixed version, and why it matters to your stack would make it something you can actually trust
Security had a rough week 😭 https://t.co/ozxwjSpABv by @sazabi is a free vuln feed with RSS, JSON and an MCP server. Upload a lockfile and you only see what hits your exact versions. No account needed. If you're into AI security too, let's connect https://t.co/ozxwjSpABv
@tirupathipodugu Exactly, and the 77% reusing its arguments is the part I keep thinking about. You're not just agreeing, you're adopting its reasoning as your own. Writing my own call down first seems like the cheapest fix
New Anthropic study: 50 pro analysts worked with Claude Code, and the agent's lean on each choice was secretly set by coin flip. They went with it 86% of the time, and 77% reused its arguments. My takeaway: make your own call before the agent weighs in. https://t.co/aIXJmokvGn
Building with multiple agents? Anthropic's free single-agent-loop-kit is a Claude Code skill that helps you fold subagents into one agent loop with skills and tools. It scores every change on an eval, so a subagent only goes when the numbers say so. https://t.co/h6d9HadT7U
@TarekCompi@sazabi Right? Checking the CVE List, KEV, EPSS and OSV separately gets old fast. Having exploitation data in the same entry as the severity is the part I like most
Free CVE feed built for agents. No login, no signup, no API key. https://t.co/ozxwjSpABv from @sazabi merges the CVE List, CISA KEV, EPSS and OSV into one entry per vuln, and has an MCP server so your coding agent can check your lockfile. https://t.co/ozxwjSpABv
@BatsouElef Thank you! Right now I'm testing a mix of AI tool finds and AI security stories to see what people actually engage with. Any one thing you'd focus on early?
@TheXSami Same, that's the part that sold me. A raw CVE feed is mostly noise until it's filtered to what you actually run. Would you wire it into Codex so it checks before you add a new package?
@bitnxs@sazabi Good question. The feed flags entries with "Fix available" and the MCP has a check_stack tool for lockfiles, but I haven't confirmed yet whether it returns the exact fixed version. Testing that next