FortiOS Local LLM Signature Library

§1

Complete Signature Index

38 signatures · 4 categories

#	Signature Name	Category	Match Field	Pattern	Weight	Purpose
01	LM.Studio.Native.API	Infra	URI	/api/v0/	50	LM Studio native REST API v0 — unique, no cloud overlap
01b	LM.Studio.Native.APIv1	Infra	URI+Header	/api/v1/ + node	50	LM Studio native REST API v1 — anchored by node User-Agent
02	AnythingLLM.API	Infra	URI	/api/v1/workspace	50	AnythingLLM server endpoint
03	Local.LLM.OpenAI.Compat	Infra	URI	/v1/chat/completions	30	Any OpenAI-format local LLM server
04	AnythingLLM.OpenAI.SDK	Client	Header	OpenAI/JS	55	AnythingLLM via OpenAI JS SDK (chat POST flows)
04b	AnythingLLM.NativeAPI	Client	URI+Header×2	/api/v + node + sec-fetch-mode	53	AnythingLLM model enumeration via LM Studio native API
05	Local.LLM.Anthropic.Compat	Infra	URI	/v1/messages	25	Any Anthropic-format local LLM server
06	LM.Studio.Anthropic.API	Infra	URI+Header	/v1/messages + claude-cli	55	LM Studio Anthropic endpoint (Claude Code path)
07	Client.ClaudeCode	Client	Header	claude-cli	55	Claude Code CLI client
08	Client.OpenAI.Python.SDK	Client	Header	AsyncOpenAI/Python	50	Python openai SDK — scripted access
09	Client.CherryStudio	Client	Header	CherryStudio	55	Cherry Studio Electron client — confirmed v1.9.4
10	Model.Llama	Base Model	Body	llama-	60	Meta Llama all variants
10	Model.Mistral	Base Model	Body	mistral	60	Mistral AI family inc. Mixtral, Devstral
11	Model.Phi	Base Model	Body	phi-	60	Microsoft Phi-3, Phi-4
12	Model.Gemma	Base Model	Body	gemma-	60	Google Gemma family
13	Model.Qwen	Base Model	Body	qwen	60	Alibaba Qwen family
14	Model.DeepSeek	Base Model	Body	deepseek	65	DeepSeek AI — elevated weight for distilled models
15	Model.Nemotron	Base Model	Body	nemotron	60	NVIDIA Nemotron
16	Model.LFM	Base Model	Body	lfm	60	Liquid AI LFM series
17	Model.GLM	Base Model	Body	glm	60	Z.ai GLM series
18	Model.Granite	Base Model	Body	granite-	60	IBM Granite
19	Model.GPT-OSS	Base Model	Body	gpt-oss	60	OpenAI open source models
20	Model.OLMo	Base Model	Body	olmo	60	Allen AI OLMo / olmOCR
21	Model.Ernie	Base Model	Body	ernie-	60	Baidu Ernie
22	Model.MiniMax	Base Model	Body	minimax	60	MiniMax M2
23	Model.Falcon	Base Model	Body	falcon-	60	TII Falcon
24	Model.Command	Base Model	Body	command-r	60	Cohere Command-R, Command-R+, Command-R7B
25	Model.InternLM	Base Model	Body	internlm	60	Shanghai AI Lab InternLM
26	Model.Solar	Base Model	Body	solar-	60	Upstage Solar
27	Model.Kimi	Base Model	Body	kimi	60	Moonshot AI Kimi — K2, K2.5, K2.6
28	Model.Hermes	Fine-Tune	Body	hermes	62	NousResearch Hermes 2/3/4, OpenHermes
29	Model.Dolphin	⚠ Uncensored	Body	dolphin	68	Eric Hartford Dolphin — all safety refusals removed
30	Model.Zephyr	Fine-Tune	Body	zephyr	62	HuggingFace H4 Zephyr
31	Model.OpenChat	Fine-Tune	Body	openchat	62	OpenChat Project 3.5/3.6
32	Model.Wizard	Fine-Tune	Body	wizard	62	Microsoft Research WizardLM / WizardCoder / WizardMath
33	Model.Vicuna	Fine-Tune	Body	vicuna	62	UC Berkeley LMSYS Vicuna
34	Model.Orca	Fine-Tune	Body	orca	62	Microsoft Orca-2 / OpenOrca community
35	Model.Airoboros	Fine-Tune	Body	airoboros	62	jondurbin Airoboros creative/roleplay
36	Model.Phind	Fine-Tune	Body	phind	62	Phind code-specialized CodeLlama

§2

Weight Hierarchy

Priority (highest wins):

68 Dolphin — uncensored

65 DeepSeek — distilled priority

62 Fine-tune orgs

60 Base model families

55 Client signatures

53 AnythingLLM.NativeAPI

50 App/server signatures

30 OpenAI catch-all

25 Anthropic catch-all

ℹ

How weights interact: When multiple signatures match the same session, FortiOS selects the highest weight. Fine-tuner signatures (62) outrank base model signatures (60) — nous-hermes-2-llama-3.1-8b correctly identifies as Model.Hermes, not Model.Llama. Client signatures (weight 55) yield to base model signatures (weight 60) on the same flow — when a model is identified, the model wins. Client signatures appear in logs for flows where no model signature fires (e.g., an unrecognized model name). AnythingLLM.NativeAPI (weight 53) sits between client and infrastructure sigs — it wins over LM.Studio.Native.APIv1 (50) on model enumeration flows where AnythingLLM's Node.js fetch is the client, while LM Studio's own http-module polling (no sec-fetch-mode) still correctly fires at weight 50. Infrastructure signatures (URI-based, weight 50 or lower) operate on distinct request types and do not compete with model signatures.

§3

Important Notes

⚠

SECURITY — Dolphin (weight 68): The Dolphin series explicitly removes all AI content safety refusals. Users running Dolphin locally bypass AI content guardrails entirely. Recommended as the first signature promoted from Monitor to Block.

⚡

Body Matching Limitation: The IPS engine does not reassemble HTTP body data across TCP packets. The "model" field appears at the start of every OpenAI-compatible POST body and is reliably in the first packet for normal requests. Verify with packet capture if matching issues arise.

⚡

Claude Code uses Anthropic format, not OpenAI format: Claude Code sends requests to /v1/messages, not /v1/chat/completions. Per-model signatures (09–35) only match on /v1/chat/completions. Use Client.ClaudeCode (#07) and LM.Studio.Anthropic.API (#06) to detect Claude Code traffic.

✓

Deep Packet Inspection required for HTTPS: Signatures using --service HTTP automatically cover both HTTP and HTTPS. HTTPS inspection requires an SSL deep inspection profile on the firewall policy.

ℹ

FortiOS 7.6.4 minimum required for Category 36 (GenAI): Category 36 was introduced in FortiOS 7.6.4. These signatures have been tested and confirmed on 7.6.6. If you are running 7.6.3 or earlier, Category 36 will not be available — substitute the most appropriate category available in your version. The F-SBID signature syntax itself works on earlier FortiOS versions.

ℹ

FortiGuard native signatures: FortiGuard already covers Ollama, Claude (cloud), ChatGPT, and other cloud AI services. These signatures cover the same API formats on local/internal network traffic. They complement — not replace — FortiGuard signatures.

ℹ

Application and Filter Overrides — per-application only: After saving signatures, add them to your Application Control profile under Security Profiles → Application Control → Application and Filter Overrides. Overrides operate at the per-application level — multi-select individual signatures and assign an action (Monitor or Block). The GenAI category (36) is already present in the global Application Control settings and does not need to be added as a separate override entry.

⚡

Layer 2 traffic — micro-segmentation required: These signatures only inspect traffic that crosses the FortiGate at layer 3. If the inference server and the client are on the same subnet, their traffic will not pass through the FortiGate and these signatures will not fire. Micro-segmentation — placing inference servers on a dedicated VLAN that routes through the FortiGate — is required to enforce inspection on same-subnet LLM traffic.

§4

CLI — Infrastructure & Client Signatures

Signatures 01–08

OpenAI-Compatible API Format (01–04)

# Paste into FortiOS CLI
config application custom

  edit "LM.Studio.Native.API"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "LM Studio native REST API v0 path - /api/v0/ is unique to LM Studio, confirmed no overlap with cloud AI services"
    set signature "F-SBID( --name \"LM.Studio.Native.API\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/api/v0/\"; --no_case; --context uri; --weight 50;)"
  next

  edit "LM.Studio.Native.APIv1"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "LM Studio native REST API v1 path - URI + node User-Agent distinguishes from cloud services. False positive confirmed on chat.qwen.ai."
    set signature "F-SBID( --name \"LM.Studio.Native.APIv1\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/api/v1/\"; --no_case; --context uri; --pattern \"node\"; --no_case; --context header; --weight 50;)"
  next

  edit "AnythingLLM.API"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "AnythingLLM server - /api/v1/workspace path unique to AnythingLLM"
    set signature "F-SBID( --name \"AnythingLLM.API\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/api/v1/workspace\"; --no_case; --context uri; --weight 50;)"
  next

  edit "Local.LLM.OpenAI.Compat"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Catch-all for OpenAI-compatible /v1/chat/completions - covers LM Studio, Ollama, llama.cpp, vLLM, LocalAI, Jan and future servers"
    set signature "F-SBID( --name \"Local.LLM.OpenAI.Compat\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --no_case; --context uri; --weight 30;)"
  next

  edit "AnythingLLM.OpenAI.SDK"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "AnythingLLM client - identified by OpenAI/JS User-Agent sent by the OpenAI JavaScript SDK"
    set signature "F-SBID( --name \"AnythingLLM.OpenAI.SDK\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"OpenAI/JS\"; --no_case; --context header; --weight 55;)"
  next

  edit "AnythingLLM.NativeAPI"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "AnythingLLM native API - node+cors vs LM Studio polling"
    set signature "F-SBID( --name \"AnythingLLM.NativeAPI\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/api/v\"; --no_case; --context uri; --pattern \"node\"; --no_case; --context header; --pattern \"sec-fetch-mode\"; --no_case; --context header; --weight 53;)"
  next

end

Anthropic-Compatible API Format + Client Identification (05–08)

config application custom

  edit "Local.LLM.Anthropic.Compat"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Catch-all for Anthropic Messages API format on local/internal traffic - covers current and future servers implementing /v1/messages"
    set signature "F-SBID( --name \"Local.LLM.Anthropic.Compat\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/messages\"; --context uri; --weight 25;)"
  next

  edit "LM.Studio.Anthropic.API"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "LM Studio Anthropic-compatible endpoint accessed by Claude Code - introduced in LM Studio 0.4.1"
    set signature "F-SBID( --name \"LM.Studio.Anthropic.API\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/messages\"; --context uri; --pattern \"claude-cli\"; --no_case; --context header; --weight 55;)"
  next

  edit "Client.ClaudeCode"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Claude Code CLI client - User-Agent claude-cli/x.x.x (external, cli) - fires whether connecting to Anthropic cloud or local server"
    set signature "F-SBID( --name \"Client.ClaudeCode\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"claude-cli\"; --no_case; --context header; --weight 55;)"
  next

  edit "Client.OpenAI.Python.SDK"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Python OpenAI SDK client - AsyncOpenAI/Python user-agent indicates scripted or programmatic LLM API access"
    set signature "F-SBID( --name \"Client.OpenAI.Python.SDK\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"AsyncOpenAI/Python\"; --no_case; --context header; --weight 50;)"
  next

  edit "Client.CherryStudio"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Cherry Studio Electron client - CherryStudio UA"
    set signature "F-SBID( --name \"Client.CherryStudio\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"CherryStudio\"; --no_case; --context header; --weight 55;)"
  next

end

§5

CLI — Base Model Families

Signatures 09–26 · weight 60 (DeepSeek: 65)

# Match on /v1/chat/completions URI + model name in POST body
config application custom

  edit "Model.Llama"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Meta Llama family - llama- avoids body text FP"
    set signature "F-SBID( --name \"Model.Llama\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"llama-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Mistral"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Mistral AI model family - covers Mistral, Mixtral, Ministral, Magistral, Devstral"
    set signature "F-SBID( --name \"Model.Mistral\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"mistral\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Phi"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Microsoft Phi family - phi- avoids body text FP"
    set signature "F-SBID( --name \"Model.Phi\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"phi-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Gemma"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Google Gemma family - gemma- avoids body text FP"
    set signature "F-SBID( --name \"Model.Gemma\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"gemma-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Qwen"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Alibaba Qwen model family"
    set signature "F-SBID( --name \"Model.Qwen\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"qwen\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.DeepSeek"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "DeepSeek AI model family - weight 65 wins over Llama for deepseek-r1-distill-llama variants"
    set signature "F-SBID( --name \"Model.DeepSeek\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"deepseek\"; --no_case; --context body; --weight 65;)"
  next

  edit "Model.Nemotron"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "NVIDIA Nemotron model family"
    set signature "F-SBID( --name \"Model.Nemotron\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"nemotron\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.LFM"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Liquid AI LFM model family"
    set signature "F-SBID( --name \"Model.LFM\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"lfm\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.GLM"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Z.ai GLM model family"
    set signature "F-SBID( --name \"Model.GLM\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"glm\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Granite"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "IBM Granite family - granite- avoids body text FP"
    set signature "F-SBID( --name \"Model.Granite\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"granite-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.GPT-OSS"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "OpenAI gpt-oss open source model family"
    set signature "F-SBID( --name \"Model.GPT-OSS\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"gpt-oss\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.OLMo"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Allen AI OLMo model family"
    set signature "F-SBID( --name \"Model.OLMo\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"olmo\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Ernie"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Baidu Ernie family - ernie- avoids body text FP"
    set signature "F-SBID( --name \"Model.Ernie\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"ernie-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.MiniMax"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "MiniMax model family"
    set signature "F-SBID( --name \"Model.MiniMax\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"minimax\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Falcon"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "TII Falcon family - falcon- avoids body text FP"
    set signature "F-SBID( --name \"Model.Falcon\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"falcon-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Command"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Cohere Command-R family - command-r pattern avoids body text FP"
    set signature "F-SBID( --name \"Model.Command\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"command-r\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.InternLM"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Shanghai AI Lab InternLM model family"
    set signature "F-SBID( --name \"Model.InternLM\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"internlm\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Solar"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Upstage Solar family - solar- avoids body text FP"
    set signature "F-SBID( --name \"Model.Solar\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"solar-\"; --no_case; --context body; --weight 60;)"
  next

  edit "Model.Kimi"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Moonshot AI Kimi model family - K2, K2.5, K2.6"
    set signature "F-SBID( --name \"Model.Kimi\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"kimi\"; --no_case; --context body; --weight 60;)"
  next

end

§6

CLI — Fine-Tune Organizations

Signatures 28–36 · weight 62 (Dolphin: 68)

config application custom

  edit "Model.Hermes"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "NousResearch Hermes fine-tune family - Hermes 2/3/4 and OpenHermes on any base model"
    set signature "F-SBID( --name \"Model.Hermes\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"hermes\"; --no_case; --context body; --weight 62;)"
  next

    # SECURITY PRIORITY — weight 68, always wins. Recommend Block after monitor phase.
  edit "Model.Dolphin"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Eric Hartford Dolphin - UNCENSORED. All AI safety refusals removed. Block recommended after monitor phase."
    set signature "F-SBID( --name \"Model.Dolphin\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"dolphin\"; --no_case; --context body; --weight 68;)"
  next

  edit "Model.Zephyr"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "HuggingFace H4 Zephyr DPO fine-tune series"
    set signature "F-SBID( --name \"Model.Zephyr\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"zephyr\"; --no_case; --context body; --weight 62;)"
  next

  edit "Model.OpenChat"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "OpenChat project fine-tune series"
    set signature "F-SBID( --name \"Model.OpenChat\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"openchat\"; --no_case; --context body; --weight 62;)"
  next

  edit "Model.Wizard"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Microsoft Research WizardLM family - WizardLM, WizardCoder, WizardMath"
    set signature "F-SBID( --name \"Model.Wizard\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"wizard\"; --no_case; --context body; --weight 62;)"
  next

  edit "Model.Vicuna"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "UC Berkeley LMSYS Vicuna Llama fine-tune series"
    set signature "F-SBID( --name \"Model.Vicuna\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"vicuna\"; --no_case; --context body; --weight 62;)"
  next

  edit "Model.Orca"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Microsoft Orca and community OpenOrca fine-tune series"
    set signature "F-SBID( --name \"Model.Orca\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"orca\"; --no_case; --context body; --weight 62;)"
  next

  edit "Model.Airoboros"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "jondurbin Airoboros creative and roleplay fine-tune series"
    set signature "F-SBID( --name \"Model.Airoboros\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"airoboros\"; --no_case; --context body; --weight 62;)"
  next

  edit "Model.Phind"
    set category 36
    set technology 2
    set behavior 9
    set vendor 0
    set protocol "TCP"
    set comment "Phind code-specialized CodeLlama fine-tune series"
    set signature "F-SBID( --name \"Model.Phind\"; --service HTTP; --protocol tcp; --flow from_client; --pattern \"/v1/chat/completions\"; --context uri; --pattern \"phind\"; --no_case; --context body; --weight 62;)"
  next

end

§7

Priority Resolution — Common Overlap Scenarios

Example Model ID	Patterns Match	Winning Signature	Reason
deepseek-r1-distill-llama-8b	deepseek (65), llama (60)	Model.DeepSeek	Weight 65 beats 60
nous-hermes-2-llama-3.1-8b	hermes (62), llama (60)	Model.Hermes	Weight 62 beats 60
dolphin-2.9-llama3-8b	dolphin (68), llama (60)	Model.Dolphin	Weight 68 always wins
dolphin-2.5-mixtral-8x7b	dolphin (68), mistral (60)	Model.Dolphin	Weight 68 always wins
openhermes-2.5-mistral-7b	hermes (62), mistral (60)	Model.Hermes	Weight 62 beats 60
phind-codellama-34b-v2	phind (62), llama (60)	Model.Phind	Weight 62 beats 60
devstral-small-2505	mistral (60)	Model.Mistral	"stral" suffix caught by mistral pattern
nous-capybara-34b	none	No model match	Falls through to infrastructure sig only

§8

Traffic Coverage Matrix

Traffic Scenario	Infra Sig	Client Sig	Model Sig	FortiGuard
AnythingLLM → LM Studio (HTTP)	✓ #01 #03	✓ #04	✓ #09–35	—
AnythingLLM → LM Studio (HTTPS+DPI)	✓ #01 #03	✓ #04	✓ #09–35	—
Claude Code → LM Studio (/v1/messages)	✓ #05 #06	✓ #07	— (Anthropic format)	✓ Cloud path
Claude Code → Anthropic cloud	—	✓ #07	—	✓ Native
Python script → local LLM	✓ #03	✓ #08	✓ #09–35	—
Open WebUI → local LLM	✓ #03	— (generic UA)	✓ #09–35	—
curl / manual API call	✓ #03	— (no UA)	✓ #09–35	—
Any client → unknown future /v1/chat server	✓ #03	✓ if known UA	✓ #09–35	—
Any client → unknown future /v1/messages server	✓ #05	✓ if known UA	— (Anthropic format)	—
Browser → ChatGPT / Claude.ai	—	—	—	✓ Native
Any client → Ollama	✓ #03	✓ if known UA	✓ #09–35	✓ Native
Client + server on same subnet (no micro-seg)	—	—	—	—

✓ = matched | — = not matched | ✓ partial = matched only if that specific condition is present

§9

FortiGuard Integration Notes

✓

These signatures complement FortiGuard — they do not replace it. The distinction is not simply cloud versus local. FortiGuard signatures may exist for some of these applications but may not match against the specific local communication mechanisms documented here — the URI paths, User-Agent headers, and POST body patterns used by local inference servers and their clients. FortiGuard signatures may also be updated in the future to detect some of these patterns, or local applications may evolve in ways that cause FortiGuard signatures to trigger where they previously did not. This library's value is in the specific identification mechanisms discovered through direct traffic analysis of local LLM infrastructure. Always verify current FortiGuard coverage in your own environment.

Application	FortiGuard Coverage	This Library Adds
Claude / Anthropic	Signatures exist targeting Anthropic API patterns	Local /v1/messages servers and Claude Code client UA not matched by FortiGuard cloud signatures
Ollama	Signatures exist for Ollama	Local Ollama API /v1/chat/completions serving — FortiGuard may not match all local serving patterns
ChatGPT / OpenAI	Signatures exist targeting OpenAI API patterns	Local OpenAI-compatible servers and Python SDK client UA
LM Studio	No FortiGuard signature (local-only app at time of testing)	Full coverage — native API, Anthropic API, model identity
AnythingLLM	No FortiGuard signature (self-hosted at time of testing)	Full coverage — server endpoint and client UA
Model families	Not applicable — FortiGuard does not identify model-level traffic	All 27 model signatures for local deployment

Verification Commands

# Verify all custom signatures saved correctly
show application custom

# Check a specific signature
show application custom "Model.Dolphin"

# Verify AIAP GenAI database is current
diagnose autoupdate versions | grep -A 6 GenAI

# View application control log entries
execute log display

§10

Network Architecture Considerations

Layer 2 Traffic and Micro-Segmentation

⚡

Same-subnet traffic bypasses the FortiGate entirely. These signatures only inspect traffic that crosses the FortiGate at layer 3. If the LM Studio server (or any other inference server) and the client making requests sit on the same subnet, their traffic never crosses a layer 3 boundary and the FortiGate will not see it — the signatures will not fire regardless of how the Application Control profile is configured.

To enforce application control on same-subnet LLM traffic, micro-segmentation is required. This means placing inference servers on a dedicated VLAN or subnet that routes through the FortiGate, so that all client traffic — regardless of client location — must cross a layer 3 boundary to reach the inference server.

Scenario	Traffic visible to FortiGate?	Signatures fire?
Client and server on different subnets (routed through FortiGate)	✓ Yes	✓ Yes
Client and server on same subnet, inference server on dedicated VLAN (micro-segmented)	✓ Yes	✓ Yes
Client and server on same flat subnet, no micro-segmentation	— No	— No
Localhost only (LM Studio and client on same machine)	— No	— No (endpoint control required)

Localhost traffic (where both the inference server and the client run on the same machine) is never visible to the FortiGate and requires endpoint-level controls rather than network-level inspection.

§11

Versioning & Update Strategy

Version History

Version	Date	Summary
v1.3.0	2026-05-06	Added `Model.Kimi` — Moonshot AI Kimi model family (K2, K2.5, K2.6), pattern `kimi`, weight 60. Total: 39 signatures.
v1.2.0	2026-05-06	Added `AnythingLLM.NativeAPI` (sig 04b, weight 53) — triple-condition signature identifying AnythingLLM model enumeration via LM Studio native API. Differentiates Node.js 18+ fetch (`sec-fetch-mode` present) from LM Studio's http-module polling (`sec-fetch-mode` absent). Fixed `Model.Solar` false positive (production-confirmed: word "solar" in chat content triggered the sig — pattern tightened to `solar-`). Proactive hyphen anchoring applied to 6 base model signatures (`llama-`, `phi-`, `gemma-`, `granite-`, `falcon-`, `ernie-`) to reduce body-text false positive risk. Total: 38 signatures.
v1.1.0	2026-05-06	Added `Client.CherryStudio` — Cherry Studio Electron client confirmed in traffic analysis. Fixed `Model.Command` false positive (pattern `command` → `command-r`, confirmed in production logs). Fixed `AnythingLLM.OpenAI.SDK` weight 60 → 55 to restore correct weight hierarchy. Total: 37 signatures.
v1.0.1	2026-05-02	Fixed `LM.Studio.Native.API` false positive on cloud AI web frontends (chat.qwen.ai confirmed). Split into `LM.Studio.Native.API` (v0 path, unique) and `LM.Studio.Native.APIv1` (v1 path + node User-Agent). Total: 36 signatures.
v1.0.0	2026-04-26	Initial release. 35 signatures covering infrastructure, base model families, and fine-tune organizations across OpenAI-compatible and Anthropic-compatible API formats.

When to Update This Library

The LLM ecosystem moves quickly. The following events should trigger a review and potential update of these signatures:

Trigger	Action	Priority
New model family released	Add base model signature (weight 60) if model name is distinct and likely to be deployed locally	Medium
New fine-tune organization emerges	Add fine-tuner signature (weight 62) if org has a consistent naming pattern across their catalog	Medium
False positive confirmed in logs	Tighten the pattern or add a secondary discriminator (URI + header combination). Document in CHANGELOG.	High
New local inference server released	Check URI paths and User-Agent strings; add infrastructure signature if unique patterns exist	Medium
New LLM client tool released	Check User-Agent header; add client identification signature if string is stable and distinctive	Medium
LM Studio version update	Verify `/api/v0/` and `/api/v1/` paths still apply; check for new API versions introduced	Periodic
FortiGuard signature update	Re-test cloud AI web frontends to verify no new false positives introduced by FortiGuard changes	Periodic
FortiOS upgrade	Re-verify all signatures save correctly and Category 36 remains available on the new version	On upgrade

Recommended Review Sources

ℹ

Monitor these sources to identify when new signatures may be needed: lmstudio.ai/models for new model additions to the LM Studio catalog, huggingface.co/models sorted by downloads for trending locally-deployable models, and the GitHub repository Issues for community-reported false positives and signature requests.

Versioning Convention

This library follows semantic versioning:

Version Type	When Used	Example
Patch (x.x.N)	Bug fixes, false positive corrections, comment updates	1.0.0 → 1.0.1
Minor (x.N.0)	New signatures added, no breaking changes	1.0.x → 1.1.0
Major (N.0.0)	Signatures removed, renamed, or weights significantly restructured	1.x.x → 2.0.0

When updating to a new version, always check the CHANGELOG for any signatures that have been removed or renamed before pasting new CLI blocks, as the FortiGate will error if you attempt to set a pattern on an existing signature without first deleting it.