■ SECURITY & SOVEREIGNTY // POLICY MATRIX
Autonomy without unbounded authority
Agentic engineering is powerful precisely because agents can act. Permissions, sandboxing, secrets, network access, approvals and verification belong in the architecture — not as post-launch hardening.
Repository read
filesystem.readrisk · low
Scoped read-only context required for reasoning and AST indexing within permitted repository paths.
Mitigation: Path normalization prevents traversal outside task workspace.
Workspace write
filesystem.writerisk · low
Required for synthesizing code diffs inside isolated ephemeral workspace or git worktree.
Mitigation: Restricted strictly to active git worktree; cannot touch root dotfiles or .git/config.
External network
http.requestrisk · high
Potential data egress channel. Agents may need to pull documentation, but unauthorized outbound requests risk exfiltration.
Mitigation: Domain allowlisting + operator confirmation dialog specifying destination URL.
Git commit & push
git.pushrisk · medium
Modifies upstream remote history and triggers external CI/CD pipelines and webhooks.
Mitigation: Operator must visually review full unified diff and sign commit approval.
Production deploy
deploy.productionrisk · critical
Direct operational impact on live customer infrastructure and production services.
Mitigation: Multi-party approval gate + dry-run verification run required.
Database schema alter
db.migrate / db.droprisk · critical
Risk of catastrophic data loss or table locks in shared databases.
Mitigation: Disallowed on production databases; permitted on ephemeral test containers with confirmation.
Shell arbitrary eval
shell.exec("rm -rf ...")risk · critical
Arbitrary shell execution with raw parameters bypasses structured tool boundary contracts.
Mitigation: Deny-first posture. Only parameterized, audited commands (e.g. npm test, cargo build) permitted.
Secrets exposure
env.dump / cat .envrisk · critical
Direct compromise of infrastructure credentials, API keys, or private SSH keys.
Mitigation: Secrets broker intercepts and masks keys before tokens enter LLM prompt context.
Unknown destructive
*risk · critical
Any undeclared, non-schema-validated tool invocation is inherently untrusted.
Mitigation: Default-deny architecture: capabilities must be explicitly declared and granted.
Default-deny by construction
Any undeclared, non-schema-validated tool invocation is inherently untrusted and is denied. Capabilities must be explicitly declared, validated, and granted — the same truth boundary that keeps UNKNOWN, UNAVAILABLE and BLOCKED states honest in the product UI.