Krypteia Sec
Jason Jeske · Agentic AI Engineer, Offensive Security
Build the attacker before someone else does.
Traditional scanners fuzz bytes. AI vulnerabilities live in meaning. Prompt injection is not a malformed HTTP request. It is a sentence that sounds helpful but changes the model's behavior. The only thing that can probe meaning at scale is another AI.
Same prompt, different response every time. Static test suites are useless. Evaluating an AI system means running hundreds of variations and measuring success probabilistically, not checking a single pass or fail.
The most dangerous AI attacks are multi-turn conversations that gradually shift context, build trust, and bypass guardrails that catch single-prompt attempts. Only AI maintains that strategic state.
A jailbroken chatbot says something inappropriate. A jailbroken agent with database access exfiltrates data, modifies records, and executes code. The blast radius is orders of magnitude larger.
Everything here is free to read: in-depth guides, a structured course, and original research on AI and agentic security.
Research on how LLMs, agents, and MCP servers actually fail. Agent-drafted, human-verified, every claim sourced.
In-depth pillar guides on MCP security, prompt injection testing, agentic AI red teaming, and AI agent security, plus a 30-lesson structured course from foundations through advanced.
Long-form articles on AI and agentic security, including a build-oriented series that walks through constructing an agentic red team from scratch.
A curated hub of AI education video content and courses from across the field, organized so the good material is easy to find.
Anthropic's new service sends vulnerability reports without human review. The useful test is how much work it takes to turn those reports into verified fixes.
OpenAI's evaluation model escaped a sandbox and reached Hugging Face production, Anthropic's reached three real companies, and a Chinese actor drove DeepSeek through Hermes at 460 targets. Not one of them needed a novel exploit.
OpenAI said on August 7 it cannot rule out Critical cyber capability in Astra, and on August 10 shipped GPT-5.6-Cyber at 95% completion on exploit-chain and privilege-escalation requests through sixteen resellers. Progress LoadMaster CVE-2026-8037 reached KEV as a pre-auth RCE at CVSS 9.6, and Exfil
In-depth guides on the attack surfaces that matter, original research, and a structured course that runs from foundations to advanced. All of it free to read.
