▸ RESEARCH

Highflame Research.

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

▸ PUBLISHED RESEARCH

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.

Pulse · arXiv:2506.07330

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.

Read on arXiv →

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 · arXiv:2602.16935

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.

Read on arXiv →

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

Whitepapers

Long-form technical papers on the architecture behind the fabric. Free to read, no gate.

▸ RESEARCH NOTES

From the research blog

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

RESEARCHMay 19, 202612 min

Mission Drift: Why AI Agents Fail at Step 100

Description: AI agents do not always fail with a crash. They drift. Learn why Step 1 testing and passive observability cannot stop Mission Drift, and how Highflame Compass provides runtime enforcement to keep autonomous agents aligned through Step 100.

Read the paper →