$ the-wire · showcase
Docker sandbox stops unpacking agent archives on your host
By RepoJournal · Filed · About Safety Evals · Composed from the cited sources · methodology
inspect_ai closed a host-side extraction path in its Docker sandbox while it opened bridge response rewriting to operators, and inspect_evals consolidated code extraction across four benchmarks.
fix(sandbox): Docker `read_file()` no longer unpacks agent-controlled archives on the host or fills its disk (#5829) UKGovernmentBEIS/inspect_ai
Docker read_file() used to let Compose extract the container-built tar archive into a host temp directory; it now streams `docker compose cp ... -` and reads the single regular file in memory, so an agent-controlled archive never touches your host disk.
feat: let agent bridges transform model responses with `response_filter` (#5715) UKGovernmentBEIS/inspect_ai
Agent bridges gained a `response_filter` hook on bridge_generate, letting you rewrite or replace model responses in place, with the compaction baseline recorded before the filter runs and a filter exception failing the sample.
fix: prevent StoreModel validation leaks and make store and instance read-only (#5748) UKGovernmentBEIS/inspect_ai
StoreModel validation used to write a throwaway model's fields into the current sample's store, or the default store, on every assignment and dump; validation no longer leaks into the ambient store, and the store and instance are read-only.
feat(agent-bridge): forward allowed Anthropic betas, Accept-Encoding and Responses reasoning context (#5708) UKGovernmentBEIS/inspect_ai
The sandbox agent-bridge proxy now forwards only allowlisted client headers (Accept-Encoding and anthropic-beta) plus Responses reasoning context, so opaque reasoning options survive the hop without passing arbitrary headers through.
feat(humaneval): use the shared extract_code_block helper (part of #2454) (#2519) UKGovernmentBEIS/inspect_evals
humaneval, bigcodebench, apps and usaco moved to the shared extract_code_block helper, so uppercase or py-aliased fences, trailing whitespace and CRLF endings no longer fall through and get executed as Markdown; usaco bumps to version 3-A in its eval.yaml, so expect extraction-affected scores to shift when you re-run.
