AI Security
The threat model that arrives with a model — untrusted text becoming instructions, and everything downstream of that.
7 topics
Prompt Injection
Any text the model reads can act as instructions, and there is no reliable way to separate data from commands — which makes this a design constraint, not a bug to patch.
core25 minThe Lethal Trifecta
Private data, untrusted content, and external communication — any two are fine, all three in one agent is an exfiltration vulnerability by construction.
core15 minHandling Model Output Safely
Model output is untrusted input to everything downstream — render it, execute it, or query with it as carelessly as you would a string from a stranger and you have the same bugs.
core20 minAgent Permissions & Blast Radius
An agent acts with whatever credentials you hand it, so the security question is not whether it will misbehave but what the worst run can reach.
core20 minData Privacy & PII
Sending user data to a model provider is a processing decision with legal weight — and prompts, traces, and memory files are all places it quietly accumulates.
core20 minGuardrails & Refusals
Filters around the model that block what it should not accept or emit — useful as defence in depth, damaging when they are your only control or tuned too tight.
deep15 minSecurity Fundamentals
The handful of ideas — trust boundaries, least privilege, defence in depth — that every specific vulnerability is an instance of.
core25 minshared