Ai safety just became a gateway plugin. apache apisix 3.18.0, released august 20, buffers streaming…
ai safety just became a gateway plugin. apache apisix 3.18.0, released august 20, buffers streaming llm output and swaps flagged tokens for a provider-compatible refusal before the client sees them.
moderation moved from the model to the proxy.
Context
Apache APISIX's release post of August 20, 2026 announces 3.18.0 with a new ai-lakera-guard plugin that uses Lakera Guard's v2 API, an external service, to scan requests, responses or both. In block mode, per the post, flagged traffic is replaced with a provider-compatible refusal body, and streaming output is buffered before release so flagged tokens do not reach the client. The same release lets ai-aws-content-moderation inspect streaming responses, in a realtime mode that checks batches while streaming and can replace the remainder, or in a final-packet mode that scores the assembled output, per a merged pull request of July 30, 2026.
The note's sentence matches the Lakera plugin in block mode, which is one plugin among several and depends on Lakera's API, with fail_open controlling what happens if Lakera is unreachable. The AWS realtime mode replaces the rest of the stream after a hit and not only the flagged tokens, and final_packet is its default with check_response off by default, so the note blends the two. A new fail_mode defaults to skip for three plugins, so traffic a plugin cannot recognize passes unchecked unless fail_mode is set to error, and body-buffering plugins now default to a 64 MiB limit. No detection or efficacy rates were read. Moderation moved from the model to the proxy is the author's framing.
Watch next
- Detection accuracy of the plugins.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 22 September 2026 at 07:53 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →