We launched Kestrel this week. Itβs a local risk classifier packed into a binary under 2 MB. It outperforms GPT-5.6 on our cybersecurity risk benchmark π¦
Introducing π¦ ShellRisk-Bench - a benchmark for evaluating the cyber risk of agent tool calls.
As opposed to other benchmarks, ShellRisk-Bench is sourced from real agent trajectories and adversarial corpora, built to test guardrails at the base rates production traffic actually sees.
Here's what we found β
Introducing the Kontext SDK
π Runtime credentials for AI agents. π
Kontext gives each agent:
π One-line credential resolution: kontext.require("github", token)
π§ͺ Scoped, short-lived credentials per tool call
π§βπ» Full audit trail - every action traced to the authorizing human
Get started: app.kontext.dev