Kai Aizen
I research adversarial AI and offensive security — kernel vulnerabilities, CVE discovery, and the failure modes of large language models. I'm an independent researcher operating as SnailSploit, based in Israel. I also break AI systems for a living, which is where the interesting part starts.
My work is built on a single observation: LLMs exhibit the same trust reflexes as humans because they learned from human-generated data.
Language models were trained on text people wrote. They didn't only absorb grammar and vocabulary — they absorbed the social dynamics encoded in how we communicate. Authority, reciprocity, social proof, urgency: the levers social engineers have pulled for decades show up in AI systems because those patterns saturate the training corpus.
So social engineering and prompt injection aren't merely analogous. They're the same attack class, executed against a different substrate.
I call this inherited vulnerabilities — AI systems inherited human trust patterns along with human language. It's the thread running through everything I publish.
It also shapes the frameworks:
- AATMF tests adversarial psychology against AI systems
- SEF tests adversarial psychology against humans
- P.R.O.M.P.T applies adversarial psychology to communication itself
Three frameworks, one principle.
I split my time across two surfaces that most people treat as unrelated. On the machine side: kernel vulnerability research, coordinated disclosure, and bug bounty work across the open-source ecosystem — Go, Python, Rust, Java, container runtimes, network firmware. On the human-and-model side: AI red teaming, jailbreak methodology, and the security of agentic systems and the protocols wiring them together.
The thesis is what connects them. The reflex a pretext exploits in a person and the reflex a payload exploits in a model are, at the level that matters to an attacker, the same reflex.
- AATMF — Adversarial AI Threat Modeling Framework. Crosswalks to NIST and MITRE ATLAS.
- Adversarial Minds — book on the psychology shared between human and machine targets.
- CVEs — coordinated disclosures across kernel and open-source software. See the disclosure record →
- Published research — Hakin9 Magazine and ongoing writeups across AI security and the frameworks.
- NVD Contributor