Skip to main content

Athena · see the system

See Every Risk. Understand Every Exposure. Prove Every Decision.

Athena maps how an AI system’s identities, data, tools and models connect — what is exposed, what it can access, what is controlled, and where investigation or remediation is needed.Within an agreed, authorized scope — and what falls outside it is reported, not assumed.

Discovery capability families

Checks grouped by the question they answer.

27 scanners — 19 defender-style, 8 attacker-style — across web, API and AI-facing endpoints. Every finding is decided against a benign control: no baseline, no finding.
  • Web injectionCan input reach something it should not?
    • Reflected XSS, error-based SQL injection, command injection and template RCE
    • SSRF, in-band XXE, header injection, directory traversal and open redirect
    • Named gaps: blind, stored, out-of-band and CRLF classes
  • Discovery and routesWhat is actually there?
    • A crawl of up to 150 pages at depth 4, within a probe budget
    • Port, subdomain and endpoint discovery, and WAF detection
    • Informational results that describe the target and are never scored
  • Access and identityWho — and what — can reach what?
    • IDOR checks with a second identity in authenticated scans
    • CSRF and CORS checks, and a brute-force simulator
  • Transport and configurationWhat do the connection and headers give away?
    • TLS checks
    • Security-header and hygiene checks
  • Secrets and software compositionWhat leaks, and what ships with it?
    • Secret-leak detection
    • Software-composition checks against a small, stated table
  • APIs under abuseWhat happens when an API is pushed?
    • API fuzzing
    • Rate-limit testing
    • JSON and NoSQL operator injection
  • AI-facing endpointsCan a prompt turn the endpoint against its owner?
    • A prompt-injection scanner for AI-facing endpoints
    • Deeper agent and tool testing is Achilles’s job
  • Retest and mappingIs it really fixed, and what does it touch?
    • Remediation retest: closed, still open or inconclusive
    • Findings mapped to NIST 800-53, OWASP 2021, OWASP LLM 2025 and DoD Zero Trust — gaps, never a certificate
    • A cached pass is not a retest

Exposure and route graph

Start with what is actually there.

AI rarely runs alone. Athena discovers the endpoints and services around it, and SPINE's route map sets declared routes against inferred ones — including the ones nobody declared. Filter the map by what you care about; the same map is listed as a table below.
  • CustomersEntry point
  • Support portalEndpoint
  • API gatewayEndpoint
  • Gateway logsLogs & delivery
  • Agent serviceAgent
  • Service identityIdentity
  • Repo & pipelineLogs & delivery
  • Pipeline secretSecret
  • Model providerAttested
  • Payments APITool
  • Vector storeData
  • Second providerUndeclared route

Illustrative system · synthetic data

Coverage manifest

What was expected, observed, assessed — and what was not.

Athena reports gaps as plainly as findings: what did not run reads not measured. Named gaps today are blind, stored, out-of-band and CRLF classes. Coverage is a statement of scope, never a compliance verdict.
Coverage · assetsIllustrative
Expected46
Declared in the agreed scope.
Observed48
Found by discovery — including 2 nobody declared.
Assessed41
Tested with a result and evidence.
Inconclusive3
Tested, but the evidence does not settle it.
Not measured4
Not reached in this run, or out of reach.

Seven observed assets end this run without a result — and the report says so. Coverage fell short, so the audit reads incomplete, never passed.

Every effect, with its authority

Three calls, three boundaries.

Every outbound effect is recorded with the authority it ran under and whether policy allowed it; an effect with no authority reads unattributed. Testing records these boundaries — it does not enforce them at runtime.
  • Allowed read

    Verified
    $ support-agent → crm.read(customer: C-1042)
    Acting for
    Customer C-1042
    Expected
    Allowed — the requester's own record
    Observed
    Allowed · one record returned

    Within the requesting customer's scope.

  • Denied data access

    Verified
    $ support-agent → crm.read(customer: C-2213)
    Acting for
    Customer C-1042
    Expected
    Denied — another customer's record
    Observed
    Denied · 403 at the CRM

    The tenant boundary held for this route.

  • Held high-risk action

    Held
    $ support-agent → payments.refund(amount: $4,800)
    Acting for
    Customer C-1042
    Expected
    Held — above the approval threshold
    Observed
    Held · routed to a human approver

    The approval threshold was respected.

Illustrative call traces — synthetic data.

Differential evidence and replay

No baseline, no finding.

Every finding is decided by a differential against a benign control. Here the same request, for a record that belongs to someone else, returns that record — and the difference from the control is the finding.
  • Finding

    The test differs from its benign control — the evidence shows it.

  • Inconclusive

    Tested, but the evidence does not settle it. Never a pass.

  • Informational

    Describes the target — an open port, an endpoint. Never scored.

Control · the customer’s own invoice

Signed in as
Customer A
Requested
Invoice 1042 (A’s)
Status
200 OK
Returned
Customer A’s invoice

Test · someone else’s invoice

Signed in as
Customer A
Requested
Changed: Invoice 2213 (B’s)
Status
Changed: 200 OK — expected 403
Returned
Changed: Customer B’s invoice
VerdictFinding — the access boundary between customers is broken. Both exchanges are kept as evidence for the fix and the retest.
Negative control

The same request for the customer's own invoice returns their own record. No baseline, no finding: without this control, Athena reports nothing.

Replay

Both exchanges are re-run from the evidence pack to show the result reproduces under the recorded conditions, and again after the fix.

Evidence pack

Request, response, identity, configuration fingerprint and timestamps, signed with Ed25519 and kept with the limits of the test.

Illustrative exchange. Customers, records and identifiers are fictional.

From finding to SPINE claim

Athena finds. SPINE remembers.

Every finding becomes evidence behind a claim in SPINE, bound to the configuration it was true of — so a deployment decision always knows what it rests on.
  1. Athena

    Finding

    Decided against a benign control, with both exchanges.

  2. SPINE

    Claim

    Bound to the configuration and policy it was true of.

  3. SPINE

    Coverage

    Expected, observed and assessed — and what was not.

  4. SPINE

    Decision

    Recalculated when the evidence moves. Your team decides.

Remediation and report outputs

Evidence that survives scrutiny.

Findings ship in Ed25519-signed evidence packs with Merkle proofs, verifiable offline — proof of origin, not of correctness. Retests run against the original, and a cached pass is never counted as a fix.
  • Evidence pack

    • Both exchanges and the negative control
    • Configuration fingerprint and timestamps
    • Signed, with stated limitations

    Synthetic example

  • Mapped finding

    • Mapped to the controls it bears on
    • Dispatched to Jira, GitHub, ServiceNow or Splunk
    • Evidence of a gap, never a certificate

    Synthetic example

  • Remediation checklist

    • Enforce ownership on invoice lookups
    • Add an authorization test to the pipeline
    • Retest against the original exchange

    Synthetic example

Retest outcomes

Closed
The original failure is absent on retest.
Still open
The original failure still reproduces.
Inconclusive
The retest ran, but the evidence does not settle it.
Not a retest
A cached pass, or a scanner switched off, cannot record a fix.

Closed means the original failure is absent on retest — not that every legitimate use still works.

Mythos AI Security

See your system the way Athena does.

Start with one environment. Athena maps what your AI can reach before anything is tested.