0
Terminal actions — allow, block, rate-limit, challenge
Monolith Ward inspects live HTTP traffic inline and makes a protective decision on every single request — adaptive, audited, and running entirely inside your own network.
Static rules miss adaptive attacksInjection variants, credential abuse, scraping, and automated probing slip past signature-only web firewalls.
No real-time, explainable insightApplication teams have little visibility into the attacks being attempted against their own APIs.
False positives erode trustBrittle rule sets create operational friction and train teams to ignore alerts.
Decisions must be defensibleSecurity and compliance need verdicts that are auditable and explainable after the fact.
Data sovereignty pressureSending production traffic to a third-party cloud is increasingly hard to justify under regulation.
Someone else's cloud, your accountabilityMost WAFs are operated off-premise — yet your team owns the incident review.
Ward runs as a sidecar beside the reverse proxy you already operate. Each request gets a verdict, a reason code, and one of four terminal actions — written to an audit trail you own.
Deployed as a sidecar alongside your existing reverse proxy. Every request is evaluated in the request path.
Each decision carries a reason code and is recorded for investigation — not an opaque score.
Runs entirely inside your infrastructure. The vendor never hosts, operates, or sees a deployment.
A live decision feed, attack-surface view, flow analytics, a policy editor, and investigation views — backed by a tamper-aware audit trail across the datastore you own.
By default, no decision data, logs, or telemetry leaves your network, and there is no outbound connection to vendor infrastructure at runtime. Customer-configured integrations (SIEM, IdP) stay under your control.
Every decision maps to a reason code and a recognised taxonomy — not an opaque score.
Behavioural and anomaly analysis catch what static signatures miss.
The detection AI is built with poisoning and evasion controls, the LLM is kept out of the decision path, and autonomous agents red-team it against defined adversarial scenarios.
A mandatory observe-first progression prevents "turn it on and break production".
Every policy change is cryptographically signed and verified before it can take effect.
Fast, precise matching for the major injection classes.
Behavioural and statistical analysis of each request.
How this request fits the session and the broader traffic shape.
Policy combines the signals into one action with a reason code.
Major injection classes, matched precisely.
The classes static signatures miss.
Watches for shifts in traffic behaviour over time.
Flags improbable navigation through your API.
Spots requests statistically unlike legitimate traffic, across many dimensions.
Learns from operator feedback to sharpen accuracy over time.
Compares payloads against known-good and known-bad exemplars.
Tunes sensitivity per endpoint to suppress false positives.
Online training is bounded — drift too far from the shipped baseline and the model auto-reverts; weights are clamped, not unbounded.
Only authenticated admins submit labels, each validated against the recorded decision — no forged "block" label on an allowed request.
Cold-start gating and fail-safe-to-allow under uncertainty — an uncertain model never auto-escalates to a block.
The payload rule set is signed (Ed25519) and verified before load; an unsigned or tampered set is rejected and fails closed.
Nothing is bundled; a customer-supplied local model only writes after-the-fact incident narratives. Untrusted content is sanitised and size-bounded, with a circuit breaker — so even a successful prompt injection corrupts a narrative, not a verdict.
Autonomous AI agents run a corpus of attack scenarios — injection, initial-access, and LLM-misuse such as prompt injection and jailbreak — against a sandboxed deployment with no production credentials and no outbound network. The engine also understands AI-agent tool-call payloads, including the Model Context Protocol.
Note — MYTHOS is a structured self-assessment, not a third-party certification: continuous adversarial evidence and a coverage map, not a rubber stamp.
You don't flip a switch and start blocking production. Ward ships defaulting to the non-blocking alert-only posture.
Evaluate and record verdicts. No traffic is affected.
Confirm signal quality with no blocking. The shipped default posture.
Block, rate-limit, and challenge — rolled out gradually behind canary gates.
Stateless sidecar scales linearly with traffic across your cluster.
A single signed container for smaller footprints and edge sites.
A signed appliance for your hypervisor — including fully air-gapped environments.
A scoped shadow-mode pilot on one representative service. Observe-only, nothing blocked, and you'll see real verdicts against your own traffic within days.