Latest post
Field notes
Posts, experiments and things learned by getting them wrong first.
Featured
Use case: Base44, the door is locked. The data is outside.
From my analyses of Base44: the login stopped the user, the APIs let the data through.
Broken Access Control across three Base44 apps: unauthenticated Entity APIs, role mass assignment and anonymous LLM consumption.
Type at least two characters.
Where to start
LLM guardrails: defending AI apps without fooling yourself
Whitelists, semantic filters and LLM-as-a-judge each solve a different problem. None of them alone is enough.
Blind command injection: automating the exploit with Nuclei
No direct feedback from the server doesn't mean no feedback at all. A timing attack turns silence into a side channel.
Whitebox pentesting: a process, not a mindset
Source code access doesn't make the job easier, it changes what the job is.