Jacob Mages-Haskins

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝘁𝗵𝗲 𝗖𝗼𝗻𝘁𝗿𝗮𝘀𝘁-𝗵𝗼𝘀𝘁𝗲𝗱 𝗠𝗖𝗣 𝘀𝗲𝗿𝘃𝗲𝗿⚡ --> Connect your AI coding agent to Contrast over a secure, OAuth-protected endpoint, with no API keys to manage and nothing to install. https://lnkd.in/g2aC67ea #APISecurity #AIcoding #RuntimeSecurity

AMD's CEO Lisa Su just killed NVIDIA’s $4,000 AI box. She walked on stage, held it in one hand, and ran a 235 billion parameter model live. No data center. No cloud. No rented GPU. The secret is a single architectural trick: Unified memory💥 Instead of relying on expensive discrete graphics cards, the new AMD chip shares a 128GB memory pool between the CPU and GPU. Look at the usable memory comparison: → Nvidia RTX 4090: 24GB → Nvidia RTX 5090: 32GB → AMD's new mini PC: up to 110GB At around $3,200, it’s not a budget device. But it easily outperforms high-end Nvidia cards on massive models just by having the space to load them. The VRAM monopoly is officially cracking 👀 More about it here → https://lnkd.in/exWsjee8

The AI bubble angst seems to have calmed somewhat, but that doesn't seem justified. For people using AI developer tools, life is changing: flat-rate plans are disappearing rapidly, and paid plans are rising in price. The anecdotal numbers place it out of reach of students, hobbyists and the open source community. Only professionals with corporate accounts will get decent service. Here's a recent run through the current math, and it's still bad. https://lnkd.in/gCAUq6iS

We tested AI security scanners against enterprise code. The API cost was $315. The triage labor downstream? $128,000. That's the AI scanner tax. And most organizations are already paying it without knowing the full bill. Our research found: - 3,560 findings from a single scan. 1,000 flagged high severity. - Only 5% agreement across three scanners on the same codebase. - 49% of findings appeared in exactly one run. Non-deterministic is not a control. - AI reviewing its own findings confirmed just 1.1% as true positives. We also debated this live with Jeff Williams, Tyler Shields, Naomi Buckwalter and the rest of the Contrast team this week. Is Mythos the emergency AppSec needed, or just a marketing one? Worth watching. Full research and the webinar are linked below. Read it before your next scanner renewal.

Does it bother you that Anthropic didn't publish any details about the costs (dollars, tokens, humans) of using Mythos/Glasswing in their latest update? My understanding is that the harness is critical, and that the approach used in Glasswing is to go through every function and tell the model there's a specific vulnerability class there. Rinse-and-repeat for every permutation. We studied different AI vulnerability analysis harness strategies and how it affects the cost. No matter how you slice it, it's tough to make this make financial sense for organizations. What do you think? https://lnkd.in/dCtJvsAB https://lnkd.in/dDJnmC2i

Constraining Satan’s Deputy In their 1995 paper Programming Satan’s Computer, Ross Anderson and Roger Needham captured the security mindset in a fantastic way: “In most system engineering work, we assume that we have a computer which is more or less good and a program which is probably fairly bad. However, it may also be helpful to consider the case where the computer is thoroughly wicked... In other words, the black art of programming Satan’s computer may give insights into the more commonplace task of trying to program Murphy’s.” There are several ideas here that feel newly relevant. First, Murphy’s computer. This is where we've been for decades. Developers make inadvertent mistakes. Complexity piles up, assumptions fail, and systems break in ways nobody intended. This is relatively easy stuff... not that we are exactly winning this fight. Second, Satan’s computer. Satan is the ultimate adversarial insider. This isn't inadvertent mistakes. He's inside your code, studying your assumptions, choosing the worst input, picking the worst timing, and trying to make the system fail on purpose. Anderson and Needham were writing about crypto protocols where the network itself may be controlled by an adversary: “We have described the crypto protocols design problem as ‘programming Satan’s computer’ because a network under the control of an adversary is possibly the most obstructive computer which one could build. It may give answers which are subtly and maliciously wrong at the most inconvenient possible moment.” AI components are powerful, naive, and opaque. They're not evil, but they're like a confused deputy: limited defenses, but access, authority, and tools inside the trust boundary. Anything received by AI can trick it — instructions, documents, code, logs, tickets, emails, tool outputs, data, web pages, user requests, training data, even Morse code. And it’s not just the response you can’t trust. It’s the entire chain of things it might affect: files, processes, keys, connections, config, environments, etc... Satan’s Computer teaches ambiguity is attack surface: who sent the message, what it means, what authority it carries, and whether it can be replayed or repurposed. AI raises the stakes because it processes ambiguity as a feature. It can blur instructions, data, context, and authority — then act on the result. Its primary use case *is* the biggest danger. Ask yourself, is there any meaningful difference between bringing Satan himself into your ecosystem vs. bringing in his deputy? Consider carefully, because Satan’s deputy is already inside your code, containers, hosts, cloud, and perimeters. Satan does not need to break in...his deputy is already there. This is not an argument against AI. It's an argument against treating AI like normal software... business as usual. We need to “constrain the deputy," but we should be honest, we're at the very beginning of a long journey. https://lnkd.in/eWvNqHfz

Jeff Williams joined Jim Love on Cybersecurity Today this week and the conversation went well beyond Mythos. Jeff made an interesting point. We have been building software the same way for decades. Requirements go in, code comes out, security gets bolted on at the end. Very little learning happens in that loop. That is why SQL injection and XSS are still on the OWASP Top 10 after 25 years. Mythos automates discovery. But Jeff sees that as just the first domino. "It's not going to produce more secure code unless you fix the rest of the factory." → Discovery → Remediation → Threat modeling and security architecture → Runtime monitoring to know if defenses are holding when attackers show up Each piece has to change before you get to software that comes out secure by design. That is the opportunity AI actually represents. Listen to the full conversation https://okt.to/o7Qvf4

𝗪𝗲 𝗴𝗮𝘃𝗲 𝗖𝗹𝗮𝘂𝗱𝗲 𝗮 𝗽𝘂𝗿𝗽𝗼𝘀𝗲𝗹𝘆 𝘃𝘂𝗹𝗻𝗲𝗿𝗮𝗯𝗹𝗲 𝗮𝗽𝗽 𝗮𝗻𝗱 𝗮𝘀𝗸𝗲𝗱 𝗶𝘁 𝘁𝗼 𝗳𝗶𝗻𝗱 𝘁𝗵𝗲 𝗯𝘂𝗴𝘀. 𝙒𝙚 𝙠𝙣𝙚𝙬 𝙩𝙝𝙚𝙧𝙚 𝙬𝙚𝙧𝙚 43. 𝙄𝙩 𝙛𝙤𝙪𝙣𝙙 10. Then we attacked it. The app said, "import failed." The attack succeeded. The entire database was already being streamed to an attacker-controlled domain. 𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱 𝘂𝗻𝗱𝗲𝗿 𝘁𝗵𝗲 𝗵𝗼𝗼𝗱: → 𝗔 𝘀𝗲𝗿𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝗽𝗮𝘆𝗹𝗼𝗮𝗱 𝗶𝗻𝘀𝗶𝗱𝗲 𝗮 𝗖𝗦𝗩 triggered deserialization at the application layer → 𝗔 𝗻𝗲𝘄 𝗳𝗶𝗹𝗲 𝘄𝗮𝘀 𝗰𝗿𝗲𝗮𝘁𝗲𝗱, /etc/passwd written to it, sent out silently → WAF: nothing. EDR: nothing. 𝗔𝗽𝗽 𝗹𝗼𝗴𝘀: "𝗶𝗺𝗽𝗼𝗿𝘁 𝗳𝗮𝗶𝗹𝗲𝗱." → 𝗖𝗼𝗻𝘁𝗿𝗮𝘀𝘁 𝘀𝗮𝘄 𝗮𝗹𝗹 𝟰𝟯 𝗜𝘀𝘀𝘂𝗲𝘀 𝗮𝘁 𝗿𝘂𝗻𝘁𝗶𝗺𝗲 —> method, stack trace, exact line of custom code → One ADR protect rule. Next attempt: "𝘿𝙚𝙨𝙚𝙧𝙞𝙖𝙡𝙞𝙯𝙖𝙩𝙞𝙤𝙣 𝙖𝙩𝙩𝙖𝙘𝙠 𝙙𝙚𝙩𝙚𝙘𝙩𝙚𝙙." 𝘽𝙡𝙤𝙘𝙠𝙚𝙙. 𝙉𝙤 𝙘𝙤𝙙𝙚 𝙘𝙝𝙖𝙣𝙜𝙚. Check out Naomi Buckwalter demo below #AISecurity #ClaudeSecurity #RuntimeSecurity