Products  /  Hyphasis
HyphasisPROMPT INJECTION DEFENSE · MIDDLEWARE

Prompt injection defense that does not degrade.

Hyphasis screens every input through a mathematical decomposition pipeline. Inputs that cannot decompose are inadmissible by construction. No classifiers. No retraining. No arms race.

Request design-partner access →See how it works
SCREENING PIPELINE
INPUT
untrusted
01 · PARSE
operator-tag
02 · ROUTE
component
03 · MEASURE
admissibility
✓ ADMIT
✕ REJECT
parse-time · no model calls · every verdict carries a reason code + decomposition trace
// THE PROBLEM WITH CLASSIFIERS

The industry defends prompt injection with pattern-matching, regex filters, or secondary LLM classifiers. All three degrade: every novel attack needs an update, every update adds false positives, and every false positive erodes trust. Hyphasis replaces all three with one structural guarantee — input must decompose through a formal pipeline to be admitted. Input that does not map to the operator set has no path through the system. It is not flagged or scored. It is structurally inadmissible.

01  /  How it works
01
Route traffic through Hyphasis

A stateless HTTP service. Point LLM traffic through the endpoint before it reaches your provider. No SDK, no integration code, no app changes.

02
Decompose every input

A three-stage pipeline parses input into an operator-tagged structure, routes it through component analysis, and maps the result to a formal admissibility space.

03
Admit or reject with a reason

Admissible input proceeds to your LLM. Inadmissible input is rejected with a reason code, a quantitative admissibility score, and a causal explanation.

04
Audit everything

Every verdict is logged with its full decomposition trace, score, and tolerance settings. The Acropolis tier adds a cryptographic proof the pipeline executed correctly.

02  /  Why an injection cannot pass
No routing path

"Ignore previous instructions" targets the model's control mechanism, not the task. The pipeline has no component for meta-level model control. The input has nowhere to go.

Failed verification

The pipeline verifies seven independent properties of every input. An injection cannot satisfy origin, purpose, or temporal verification. Failure on any one is sufficient for rejection.

Quantitative distance

Even wrapped in legitimate input, a mixed payload produces a measurable gap on the injection component. Wrapping an injection in context does not eliminate the gap.

03  /  Not a classifier

Structurally different from everything else on the market.

Classifiers
Hyphasis
Novel attacks
Degrade — require retraining
Invariant — cannot decompose
False positives
Tunable, never zero
Structural — admissible always decomposes
Audit trail
Probability score
Score + full decomposition trace
Update cycle
Continuous retraining
Zero — operator set proven complete
Latency
Inference-time (model call)
Parse-time (no model call)
04  /  Who it's for

For teams deploying LLM endpoints where a prompt injection is not embarrassing — it is a compliance, safety, or security event.

AI-assisted process monitoring in energy, oil & gas, chemical operations
LLM-powered customer interfaces in healthcare & financial services
AI agents in defense & government environments
Any regulated deployment where the audit must be provable
Integration tiers
Standard
Default tolerance · 30-day audit retention · verdicts with reason codes
Enterprise
Configurable tolerance · 90-day to 1-year retention · full decomposition traces · dashboard
Acropolis
Cryptographic STARK proof per verdict, independently verifiable. For when "why was this admitted" has regulatory force.
05  /  Specs
DeploymentStateless HTTP middleware
Benchmark1,410 / 1,410 passed
Latencyp50 <10ms · p99 <50ms
Throughput>10K req/s/node
False negative rate0% structural
Update cycleZero
Live endpointapi.metacortexdynamics.com:8443
AvailabilityDesign-partner access
API SURFACE
POST /v1/screen
POST /v1/screen/batch
GET  /v1/verdict/{&lbrace;id}&rbrace;
GET  /v1/audit/stream (SSE)
PUT  /v1/config/tolerance
GET  /v1/health
BEST FOR
Energy / utilityHealthcare / life scienceDefense / governmentFinancial services
06  /  FAQ
Does Hyphasis replace my LLM?

No. It sits between user input and your provider, screening input before it reaches your model. Your LLM, prompts, and application logic are unchanged.

Does it need retraining?

No. The pipeline operates on a proven-complete operator set. Novel attacks don't require updates because novel attacks cannot decompose. There is no classifier to retrain.

What does "1,410 / 1,410" mean?

Benchmarked against 1,410 cases — legitimate inputs, known injections, novel variants, adversarial payloads. All received the correct verdict: no false negatives, no false positives.

How is this different from a regex or secondary classifier?

Regex misses novel attacks; secondary classifiers add latency and degrade. Hyphasis runs at parse-time with no model calls and a complete operator set — complete by proof, not by coverage.

Can I tune the sensitivity?

Yes — a tolerance parameter per deployment. Tighter rejects more aggressively; looser admits more but logs lower admissibility scores for review.

Is it generally available?

Hyphasis is available for design-partner access. The live endpoint is operational at api.metacortexdynamics.com:8443.

Put it in front of your endpoint.

Design-partner access is open. The endpoint is live.

Request design-partner access →All products