AI Security
LLM Output Filtering as a Security Control
Output filters are the last line before the user and the tool call. We cover when they work, when they fail, and how to measure them honestly in production.
Feb 5, 20265 min read
Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
Output filters are the last line before the user and the tool call. We cover when they work, when they fail, and how to measure them honestly in production.
A jailbreak in a model you ship downstream is a supply chain incident, not a trivia item. Here is how to reason about it and where the defensive controls belong.
Path traversal is the vulnerability class that punishes lazy analysis. Framework-specific path normalisation, OS-dependent separators, symbolic link resolution, and archive extraction all hide exploitable gaps behind code that looks defensive. Griffin's engine resolves path operations with actual semantics; Mythos reads the variable name and calls it a day.
A security AI that refuses too often is useless. One that refuses too rarely is dangerous. Griffin AI publishes calibrated refusal benchmarks; Mythos does not.
LLM spend forecasting is where finance teams meet AI engineering for the first time. The patterns that produce predictability are specific.
DeepSeek Coder has become a favourite for code-focused workloads. This is how it compares to Griffin AI when the job is security review, not code generation.
Finding a bug is not the same as proving it is exploitable. How Griffin AI synthesises concrete exploit paths and why pure-LLM scanners rarely get past the sketch stage.
Engine work parallelises cleanly. Model calls do not. We explain why Griffin AI's throughput scales with CPU while Mythos-class tools bottleneck on rate limits.
Weekly insights on software supply chain security, delivered to your inbox.