Memory Vaccination · Live

One agent is attacked. The whole fleet immunizes.

When Sgraal's detection layers catch a memory attack on one agent, it doesn't just block that agent. It extracts the attack's signature, stores it encrypted, and uses it to recognise the same attack instantly across every other agent in the fleet. A first victim becomes the whole fleet's immunity.

! Agent A — attacked Attack signature encrypted · AES-256-GCM Fleet — immunized

Detect once on Agent A → extract + encrypt the signature → the rest of the fleet recognises the same attack immediately.

The biological analogy

A vaccine works by teaching the immune system to recognise a pathogen's signature without having to survive the disease first. Herd immunity extends that protection across a population: once enough members carry the recognition, the pathogen can't spread.

Sgraal's Memory Vaccination is the same idea for an agent fleet. The first agent to encounter a novel memory attack — an injection, a poisoning pattern, a cross-agent drift cascade — is where the “infection” is recognised. Sgraal turns that single detection into a reusable signature, and every other agent inherits the recognition without having to be attacked itself.

How it works

  1. Detect. Sgraal's post-reconciliation detection layers flag a manipulated memory state on one agent (see Protect).
  2. Extract. The distinguishing pattern of that attack is distilled into a compact signature — not the underlying content, just the recognisable shape of the attack.
  3. Encrypt at rest. The signature is stored encrypted with AES-256-GCM. Signatures are protected both in transit and at rest; the recognition spreads, the raw attack material does not.
  4. Immunize. On subsequent preflights across the fleet, the stored signatures let Sgraal recognise the same attack class immediately — before it reaches a decision.

Fleet phase — for the modellers

Memory attacks spread through a fleet the way an infection spreads through a population, so Sgraal models fleet memory health with an SIR-style (susceptible–infected–recovered) analog. A fleet sits in one of three phases:

  • Sub-critical — healthy memory dominates; an attack on one agent stays contained.
  • Critical — the fleet is near the tipping point where a compromise can begin to cascade.
  • Super-critical — conditions favour spread; vaccination and quarantine matter most here.

Vaccination is what keeps a fleet sub-critical: each immunization removes susceptibility, the same way vaccination raises a population's herd-immunity threshold. The phase is computed per domain against a domain-specific critical threshold — surfaced via the fleet health-phase endpoint, with the calibrated thresholds kept server-side.

The observable surface

Vaccination happens automatically when a detection fires — there is no “upload an attack” endpoint. What you can observe and manage:

# List the vaccine signatures protecting a domain
curl -sS "https://api.sgraal.com/v1/vaccines?domain=customer_support" \
  -H "Authorization: Bearer $SGRAAL_API_KEY"

{ "domain": "customer_support", "count": 3, "vaccines": [ /* tenant-scoped */ ] }

# List agents currently flagged as compromised
curl -sS https://api.sgraal.com/v1/compromised-agents \
  -H "Authorization: Bearer $SGRAAL_API_KEY"

{ "count": 1, "agents": ["agent-117"] }

Signatures are tenant-scoped: you see and manage your own fleet's immunity. A signature can be retired with DELETE /v1/vaccines/{signature_id}, and an agent cleared with DELETE /v1/compromised-agents/{agent_id}, once remediated.

What this does not do

Cross-tenant sharing

Vaccination immunizes your fleet from your own detections — signatures are tenant-scoped and encrypted per tenant. A cross-organisation immunity network (signatures shared across tenants with a verifiable trust model) is a separate, future capability, not what this endpoint does today.

Catch a never-before-seen attack with zero cost

The first agent still has to be detected by the standard detection layers — vaccination makes the second and subsequent encounters instant. Detection effectiveness to date is validated against synthetic adversarial corpora (memory injection and cross-agent drift); production calibration is pending paying-customer onboarding. See the threat model.

Versioning

Last reviewed: 2026-05-24. Status: Live · doc-only (endpoints in production).

Related: /protect · /benchmark (memory-injection & cross-agent drift defense) · /docs/api · /docs/insights