Research · Highflame

Highflame Research

Original work on agent identity, authorization, and runtime security, adversarial findings, threat breakdowns, and the research that powers every safeguard we ship.

Peer-reviewed work behind the models

The transformer architectures powering Highflame's in-house guardrail models, published in full, with how each benchmarks against the field.

JavelinGuard: Low-Cost Transformer Architectures for LLM Security

Compact, stateless guardrail models (~110M to 450M params) for single-request analysis. Over 95% F1 on prompt injection at sub-100ms latency.

How Pulse compares: state-of-the-art accuracy at a fraction of the latency and footprint.

Model Avg latency (CPU) Size (params) Injection F1 Edge
Pulse ~47ms ~450M ~95.3% Excellent
DeBERTa-v3-base (tuned) ~80-150ms ~184M ~93-95% Excellent
LlamaGuard (7B) ~500-1000ms+ ~7B High (context-dependent) Poor
Large commercial LLM API (e.g. GPT-4) ~1000-3000ms+ Billions+ High (slow, costly) Not applicable

DeepContext: Stateful Real-Time Detection of Multi-Turn Adversarial Intent Drift in LLMs

Stateful, multi-turn guardrails that track conversational intent drift in real time, catching attacks that single-request filters miss.

How DeepContext compares: recurrent intent-tracking beats repetitive large-model inference on multi-turn attacks.

Model Multi-turn F1 Avg latency (T4 GPU / API) Defense strategy
DeepContext 0.84 ~19ms Stateful (RNN / GRU)
Granite-Guardian-3.3 (8B) 0.67 ~125ms Stateless (LLM)
GPT-5 Nano 0.63 ~317ms (API) Stateless API
LlamaGuard 4 (12B) 0.51 ~43ms Stateless (LLM)
AWS Prompt Attack Guardrails 0.38 ~235ms (API) Stateless API
Azure Prompt Shield 0.19 ~77ms (API) Stateless API

Whitepapers

Long-form technical papers on the architecture behind the fabric.

The Agent Control Fabric

The identity, policy, and enforcement substrate for production agents, and why retrofitting human IAM and API gateways into the job they were never built for does not compose. 20 pages.

Agent Identity: A Technical White Paper

The Highflame Identity layer in depth: delegated authority, scope attenuation, DPoP sender-constrained tokens, CIBA human approval, and cascade revocation. Companion paper to ACF-001.

From the research blog

Adversarial findings, threat breakdowns, and field notes from the team.

The findings ship as controls.

Every result here turns into detection and policy in the fabric, then we re-scan to prove the fix.