- Input normalised before detection, so encoded or disguised injection phrases are caught.
- Tool inputs checked before a tool runs; a person’s approval and certificate pinning for registered tool servers made available.
- Streamed replies scanned so that sensitive data split across two chunks is still caught.
What’s new
In August 2026 Kapālins strengthened the prompt-injection defences of its AI gateway, and checks for sensitive data inside AI responses while they are still streaming to the user.
From 2 August the gateway normalises input before it looks for injection, so a phrase it already recognises is also caught when it is disguised in base64 or hex, with lookalike or invisible characters, or in leetspeak. The 3 August release labelled untrusted input by its provenance, validated tool parameters before a tool received them, made a person’s approval available before chosen agent actions, flagged attempts to plant instructions in an agent’s memory, made certificate pinning available for registered MCP servers and learned a normal pattern for each agent and flagged departures from it.
Why it matters
For agent traffic that passes through the gateway, several of these defences sit where an agent acts (at the tool call, the memory write and the connection to a tool server), not only at the prompt.
Streaming needs its own control, because text already sent to the user cannot be retracted.
How it works
The gateway scans a streamed reply in a sliding window that holds back enough text that a card number or secret split across two chunks is still caught. The client receives only redacted text, and the audit record receives the original.
When an outbound rule could block the reply, or the detector cannot run, streaming is refused before the first byte and the reply is checked whole instead.
Every rule the 3 August release added started by recording only, without blocking. Each organisation decides when a rule starts to block.
Availability
The injection defences reached production on 2 and 3 August 2026, and streaming with mid-stream checks on 8 August.
AI Gateway & Prompt Firewall and AI Data Security are included in Growth, with evidence in Fortress, and are view only or an add-on on Foundation.
