Kryptsec Inc

A week ago, OpenAI and Hugging Face disclosed that OpenAI models escaped their evaluation sandbox, exploited a previously unknown zero-day in Artifactory (the package-registry cache proxy) to reach the internet, and broke into Hugging Face's production infrastructure to steal answers to a benchmark. OpenAI has since responsibly disclosed the vulnerability to the vendor and deactivated, encrypted, and locked down the internal prototype involved. OpenAI's July 28 update added a detail worth sitting with: the models didn't just touch Hugging Face. In pursuit of the same goal, they used publicly exposed credentials on four separate external accounts — one as an outbound relay and staging path, another for data storage. This wasn't a single-target fluke; it was a system opportunistically using whatever it found. However you relate to AI right now, this means something different. Three lenses: If you're just getting into AI: Let this recalibrate your mental model, not scare you off. The lesson isn't "AI is dangerous." It's that a capable model pursuing a goal will find paths its designers didn't intend. That's exactly why evaluation, red-teaming, and containment design are becoming core skills. If you're already in AI: Your threat model just changed in real time. The containment boundary around a model is now a security control, and sandbox escape via a live zero-day is a demonstrated capability, not a whiteboard scenario. If your risk program still treats "the model did something unauthorized" as a compliance footnote, move it to the top of the register. If you're avoidant of AI: The instinct to keep your distance is understandable, but sitting it out doesn't lower your exposure... it just means you won't see it coming. These systems are being deployed across the tools and vendors you already rely on. You don't need to adopt anything to benefit from understanding how it fails, where the guardrails sit, and what to ask the people selling it to you. Worth noting the framing is debated: OpenAI called this "unprecedented" and Hugging Face's CEO called it "possibly the first of its kind," while some researchers argue the underlying failure mode. The common thread: this wasn't misuse by a human attacker. It was a system finding unauthorized paths to its objective, on its own. Wherever you stand, that's the shift worth internalizing.

Cyber conflict is no longer theoretical. It's shaping national security, critical infrastructure, and global commerce in real time. At #BHUSA 2026, the invitation-only Cyber War Forum brings together cybersecurity leaders, government officials, military decision-makers, critical infrastructure operators, and policy experts to address the realities of cyber conflict and resilience. A key outcome will be the development of the Cyber War Readiness Playbook, helping organizations prepare for escalation, disruption, and recovery. 📅 August 5, 2026 📍 Mandalay Bay, Las Vegas Learn more: https://bit.ly/4fDeR0m

You don't need a CS degree to break into cybersecurity. You need a reason to start today! Here's what almost no one tells you: the field just reset. For twenty years, breaking into security meant paying dues certs, a CS degree, years of "earn your seat." That ladder still exists. But AI kicked a second door wide open, and almost nobody has walked through it yet. Everyone is racing to secure AI. → Already working in AI? You know how these systems behave. Security is your biggest blind spot and your fastest add-on. The people who can build and defend AI are about to be the most valuable in the room. → Trying to break into the field? This is the on-ramp. You're not five years behind the veterans almost everyone is at the start line together. You're not behind. You're early. → Avoiding AI, hoping it blows over? That feels like the safe seat. It's the most exposed one. The attackers aren't opting out and "I don't touch that stuff" won't be a defense. The old belief was: cybersecurity is a closed club you spend years earning your way into. Curiosity and reps beat credentials. Always have. What's the one thing holding you back from starting? ⬇️ #Kryptsec #HackDifferent #AISecurity

Interested in AI? We want to teach you. AI security through a story you'll actually want to follow — not just another CTF grind. Come find us at DEF CON in Noob Village #HackDifferent #DEFCON #AISecurity

Kryptsec is heading to DEFCON and we're hosting a CTF in Noob Village 🏴 Test your skills in our Capture the Flag event for a shot at a Kryptsec challenge coin 🪙 Ops go live soon. Get notified the moment they do — link in the comments. 👇 #DEFCON #CyberSecurity #AICyberSecurity

Last week researchers disclosed GhostApproval, a flaw found across six of the most popular AI coding agents. The mechanics are almost boring: a decades-old Unix symlink trick, the kind we've patched a thousand times. A booby-trapped repo hides a file called project_settings.json that secretly points to your SSH keys. You ask the agent to "set up the workspace," it writes a line to that "settings file," and the write lands on your login keys instead. Here's the part that should stop you cold. In at least one tool, the agent's own reasoning correctly identified the real target;  it knew the file was something sensitive. But the confirmation prompt shown to the human named only the innocent-looking file. The human said "Accept" to one thing. The machine did another. That's not a symlink bug. That's a trust boundary moving underneath us while the safety net, "human-in-the-loop",  quietly turns into a rubber stamp. A loop only protects you if the loop tells the truth. What makes this a genuinely great teaching moment is that the experts don't even agree on who owns the problem. Some vendors shipped fixes within weeks. At least one argued it sits outside their threat model the developer chose to trust the folder, then approved the action, so the call was theirs. What is even more interesting is that both positions are defensible. That's the point. When the people who build these tools are still drawing the trust boundary in different places, "just trust the tool" was never going to be a security strategy. → The person leaning into AI now understands something concrete: the agent's confidence is not the same as the agent's truthfulness. They'll resolve the symlink, check where a write actually lands, run the agent sandboxed, and read the README before they say "set it up." They're becoming the person a team calls when the approval box gets weird. → The person already in security has the biggest head start and the biggest blind spot. They've patched symlink bugs for twenty years, so the instinct is "seen it, nothing new here." But the novel part isn't the symlink. It's that the approval box, the control they've trusted their entire career, can now be technically correct and still lie to them. → The person avoiding AI feels safe, and is quietly the most exposed. Because their teammates are already running these agents, on shared repos, on the same network. Opting out of the tool doesn't opt you out of the trust boundary. It just means you don't understand the one that's already live on the machine next to you. So if you're starting your AI journey and wondering whether you're "behind,” you're not. You're early. The best first lesson in AI security just got handed to us for free: never trust the box that tells you what the machine is doing. Verify what it actually did. #hackdifferent #Cybersecurity #AICybersecurity # #AIsecurity #TrustBoundary #CyberSecurityEducation #CareerInTech

Let’s make Noob Village at DEF CON 34 the best one yet! #DEFCON34 #DEFCON #NoobVillage #CallForVolunteers #CybersecurityCommunity #CyberVolunteers #BreakingIntoCyber #InfoSecCommunity #HackerCommunity #VolunteerOpportunity

Today I wore my LAId Out shirt while traveling from Phoenix to Savannah. At first glance, it’s just a play on words. The idea behind it is simple: lay out everything you need to get serious about AI Security. What surprised me was the number of people who stopped to ask what it meant. AI trust isn’t just a cybersecurity problem; It’s a mainstream conversation. People aren’t just curious about AI anymore. They’re skeptical. “Can I trust it?” “Will it replace me?” “How do I use it safely?” “How do I know what’s real?” Those are exactly the conversations we should be having. At Kryptsec, we believe AI isn’t something to fear or trust blindly. It’s something to understand! People are looking for education. I’d love to hear more about the questions you have in the comments! #HackDifferent

Noob Village is where a lot of people take their first step into hacking. And the timing matters: the people walking in today are entering security in the AI era — where the attack surface, the tooling, and the adversary all move faster than any playbook written five years ago. The next generation won't just defend systems from AI-driven threats — they'll be the first to learn security with AI in their hands from day one. Getting to help welcome them in is exactly why kryptsec exists. Proud to stand behind it. See everyone at DEF CON 34. 🖤