Assistant: codex Assistant-Model: gpt-6-astra Assistant-Session: 01a0726e-5232-73f2-aaca-2c05ceb62efb
1.1 KiB
Credential transport acceptance
Installed sand-boxer synthetic proof 880f749e: child receives only synthetic
ANTHROPIC_API_KEY, stdout is redacted, next exec has no key, wrong-project
request refuses and workspace is removed. No real key was read.
Full checks: sand-boxer lint+171 tests, secrets-engine226 tests, Glas101 tests. Existing persisted-manager egress and deterministic Glas boundary proofs pass.
Native proposal: secrets-engine/catalog/glas-claude-agent-dev-anthropic.yaml; SECRETS-WP-0009 tracks live activation. The plan checks the existing platform mount, creates one exact data-only read policy/AppRole, TTL5m/max15m, single-use SecretID5m, token uses8. No metadata/list/write grant is included.
The real production CLI exec invocation with service-jwt selection refused:
production action 'exec' requires a durable access-engine decision record; live production remains disabled. No unsafe-demo flag, real key retrieval or
operator-token fallback was used. This is an upstream authorization/service
implementation dependency, not a successful model authentication proof.