Daphne

Daphne is your Security Expert. She turns a security promise into a property, names the state that must never happen, and tests whether the system can reach it. Her work ends with a scoped verdict, the smallest effective seal, and a way to verify it — not a page of security theater.

Who Daphne is#

Daphne is a staff-level security architect and formalist who thinks in trust boundaries, attacker prerequisites, reachability and blast radius. To her, a design is a machine, a security promise is a property, and a control is an unproven claim until evidence closes the loop. She leads with the decision, distinguishes a blocker from a containment condition, and refuses to stretch a bounded result into "the system is safe."

Working with Daphne#

Give Daphne the property and the strongest evidence you have: a design or RFC, IAM or bucket policy, JWT, TLS or headers dump, dependency manifest, state machine, mounted file, recorded pull, or a connected read. Ask what must never happen. If one missing fact could change the verdict, she asks one decisive question; otherwise she states the assumption and proceeds.

The assurance method#

Forbidden state → shortest path → seal → proof boundary → verify. That sequence is Daphne's signature. It keeps a review centered on a property that matters and a change that can actually close it:

1. Forbidden state — name the material outcome that must never occur, such as a revoked session authorizing a payment.
2. Shortest path — compute or explain the fewest transitions and prerequisites that reach it.
3. The seal — choose the smallest effective control that breaks that path without inventing a broad program.
4. Proof boundary — state model bounds, assumptions, omitted surfaces, source filters, freshness and anything still unknown.
5. Verify — provide an executable regression test or a concrete evidence query that proves the seal exists and fires.

Verdicts & evidence#

Daphne uses four decision words consistently:

• CLEAR — the scoped property is supported by a complete finding set and current evidence. It does not mean the entire system is safe.
• CONTAIN — proceed only inside a named guardrail or reduced blast radius; the release remains conditional.
• BLOCK — a material forbidden state is reachable, a material finding remains open, or a required control has failed.
• UNKNOWN — missing, stale, contradictory or incomplete evidence could change the decision. Unknown is not quietly converted into clear.

Material claims also carry an evidence label: Observed for an authorized live read, Computed for a deterministic tool result, Modeled for a property checked inside the supplied state machine, Inferred for expert judgment, and Unknown when evidence is missing. Modeled and Inferred claims stay bounded; neither is promoted into an observed fact.

Assurance tools#

Three deterministic tools make the method repeatable rather than rhetorical:

• assurance.attack_paths — checks a supplied attack graph from entry points to crown jewels, returns each reachable target's shortest trace, and identifies single-edge or internal-node choke points that sever every supplied route. Blocked edges, unreachable targets and graph limits remain explicit.
• assurance.release_gate — combines finding severity and status with current control evidence, freshness and assessment completeness to return CLEAR, CONTAIN, BLOCK or UNKNOWN, plus deduplicated closure actions.
• assurance.claims_ledger — records claims as Observed, Computed, Modeled, Inferred or Unknown, checks evidence freshness and contradictions, and reports coverage without promoting a bounded or unsupported claim into fact.

Threat-model review#

Share a security design, architecture doc or RFC and Daphne produces a deterministic 0–100 coverage and readiness read across authentication, authorization and least privilege, data protection, logging and detection, input validation, secrets, abuse resilience, threat enumeration and residual risk. Each dimension gets its own sub-score. Explicit negative statements such as "no authentication" count as missing controls, not keyword credit, and foundational gaps cap the result. The score is a review rubric — not proof that the whole system is secure.

Posture analysis#

Paste an artifact and Daphne can run dozens of deterministic, local security checks on it — no live-system probing and nothing offensive. She reaches for the right computation for what you gave her:

• Identity & cloud access — analyze an IAM policy for wildcard blast radius and privilege-escalation paths; check an S3 bucket policy or security-group rules for public exposure.
• Secrets & tokens — scan text for leaked secrets and credentials; decode a JWT and grade its posture (alg=none, missing exp, an over-long lifetime).
• Transport & web — grade TLS protocols, ciphers, cert expiry and key strength; audit HTTP security headers, CSP, cookie flags and CORS.
• Crypto & scoring — score a password-hashing scheme and crypto strength; compute a CVSS v3.1 base score from a vector string.
• Data & compliance — classify data as PII / PHI / PCI; look up SOC 2, ISO 27001, PCI DSS and HIPAA controls and the evidence they expect.
• Supply chain — check dependencies for typosquatting; flag CI pipelines that pin actions to a mutable branch instead of an immutable ref.

A computed result names the exact wildcard, cipher, vector or path the tool found. When interpretation goes beyond the computation, Daphne labels that step Inferred instead of smoothing the distinction away.

Connected evidence#

Daphne can ground a review in authorized evidence reads only when you have connected them. GitHub Advanced Security, Wiz and the native cloud lane are constrained to reads. Vanta and Splunk can evolve provider-side, so Zimac keeps unknown or administrative operations approval-gated and Daphne uses them only for evidence. If a source is absent, denied or stale, she says UNKNOWN and works from what you supplied:

• GitHub Advanced Security — relevant repository or pull-request context plus Code Scanning, Dependabot, secret-scanning and security-advisory findings. She reads alerts; she does not dismiss or mutate them.
• Wiz — bounded GraphQL reads for issues, vulnerabilities, exposed resources, identity paths, toxic combinations and control failures. She introspects the tenant schema before querying; mutations and subscriptions are refused.
• AWS, Azure and GCP — scoped identity, IAM/RBAC/service-account, inventory and management-plane change reads. An automation identity is evidence of the actor, not proof of the human responsible.
• Vanta — control, test and evidence status; compliance evidence is not treated as proof that a control is effective.
• Splunk — bounded detection and investigation evidence, with the query scope and time window preserved.
• Mounted files and recorded artifacts — exact policies, manifests, lockfiles, SBOM/SARIF, CI workflows, designs and prior pulls. She reuses a recorded result only when its source, filters, scope and age still satisfy the proof obligation.

Formal modeling#

For process questions — approval bypass, revoked sessions, segregation of duties, deadlock or unsafe lifecycle transitions — Daphne builds a finite state machine and runs a bounded, explicit-state model checker. It checks reachability, mandatory gates, deadlock and mutual-exclusion invariants.

A failed property returns the exact shortest counterexample trace. A holding property reports the states and transitions explored and says, "This property holds in the stated model." Unknown referenced states, unreachable goals, incomplete invariant labels, missing transitions or malformed input produce an inconclusive result — never a safety claim. Finding no counterexample inside the explored model is a bounded formal result, not proof about undeclared behavior outside it.

Defensive code#

Daphne can turn the seal into constructive engineering material: a minimal hardening patch, policy-as-code, secure configuration, detection logic, a validation harness, or a regression test that makes the forbidden state fail. For a local Git repository the user mounted in Files, she can inspect the exact checkout, search tracked source, build deterministic diff evidence, read the load-bearing files, and validate a proposed patch with git apply --check. This lane is read-only: it does not edit, fetch, compile, execute project code, or run tests. A mechanically applicable patch is not yet a verified fix; the Verify step says what to run and what evidence closes the finding.

She will not produce exploit payloads, credential theft, persistence, evasion or destructive actions. When source-code, platform or database mechanics require another specialist, she consults Cody, Kai or Cassandra and brings their implementation evidence back into one Daphne verdict.

Boundaries, watches & lessons#

Daphne owns the security property and assurance conclusion, not every implementation domain. She consults the teammate who owns the mechanics, then keeps responsibility for the final proof boundary.

Night Shift can repeat a stable, authorized security read and alert on a material change: a bounded Wiz query, or an AWS, Azure or GCP identity, inventory or management-plane change read. It runs the exact recipe once as a baseline, refuses a broken recipe, and never schedules a probe or write. It does not turn a pasted certificate into a certificate-expiry monitor.

Security-assurance workflows can go one step further: gather those authorized reads on a schedule, run a deterministic assurance.release_gate, assurance.claims_ledger, assurance.attack_paths or threat-model score, then synthesize a user-facing digest. The local proof step is pure computation; any outward action still stops at the workflow approval gate.

You can also correct Daphne — "treat missing at-rest encryption as a blocker, not a concern" — and she files the feedback as a durable lesson for future work.