AI overview

FreeDR has an optional AI investigator and a few on-demand AI helpers. AI is off by default. Every detection, alert, case, Hunt and response feature works without it.

When AI is on, FreeDR sends bounded, redacted case evidence to one model provider that you choose and pay for with your own API key. The model reads FreeDR evidence through a fixed list of tools. It gives a verdict that must cite retained evidence. Its tools include enrich, which runs the same threat-intel lookups (VirusTotal, AbuseIPDB, OTX and others) as case enrichment; see External enrichment.

What AI can change

The AI is not read-only. When enabled, it can:

• resolve a case when its verdict is benign;
• run four containment actions on targets it cites: terminate a process, quarantine a file, remediate (disable) persistence, block traffic. The Settings switch Let AI contain threats it finds controls this, and it is off by default: you must turn it on;
• file pending exclusions. These suppress nothing until you approve them.

It cannot isolate the host, permanently delete files, restore or release anything, approve exclusions, change built-in rules, change settings or run shell commands. For details, see AI features.

Providers and default models

You select one provider in Settings → Optional AI investigator or in the setup wizard. Each provider has a fixed HTTPS endpoint. FreeDR does not support custom endpoints, proxies, redirects or automatic failover to another provider.

Provider (Settings label)	Endpoint	Protocol	Default model ID
CheaperInference (default provider)	https://api.cheaperinference.com/v1/chat/completions	Chat Completions, bearer key	deepseek-v4.1-flash
OpenAI	https://api.openai.com/v1/responses	Responses API (store: false), bearer key	gpt-6.1-sol
Anthropic	https://api.anthropic.com/v1/messages	Messages API (x-api-key, anthropic-version: 2023-06-01)	claude-sonnet-5-5
DeepSeek	https://api.deepseek.com/chat/completions	Chat Completions, bearer key	deepseek-flash

• Model ID is editable. You can enter any model ID that your key can use. It must be 1–128 characters from letters, digits and . _ : / -. When you change the provider, FreeDR selects that provider's default model and thinking level.
• FreeDR sends x-ci-techniques: off to CheaperInference so that optional account optimizations cannot replace the model or the evidence.
• Each run records the provider and model it used. If you change the provider or model during a run, the run stops at its next round with "AI provider or model changed; start a follow-up to continue".

Thinking levels

Each provider has its own thinking (reasoning) levels. FreeDR uses the provider's names and does not map them to one shared scale. Settings lists only the levels of the selected provider, from least to most thinking.

Provider	Levels	Default	Request parameter
CheaperInference	none low medium high xhigh max	medium	reasoning_effort (none turns thinking off)
OpenAI	none low medium high xhigh max	medium	reasoning.effort
Anthropic	disabled between_tools low medium high xhigh max	medium	disabled / between_tools → thinking.type; other levels → output_config.effort (adaptive thinking)
DeepSeek	disabled low high max	high (DeepSeek has no medium)	disabled → thinking: {type: disabled}; others → thinking: {type: enabled} plus reasoning_effort

• Some levels exist only on some models. Anthropic disabled and between_tools work only on some models, and not every OpenAI model offers none. If the provider rejects a level, FreeDR shows the provider's error.
• Thinking tokens count against a reply's output limit. At any level other than none or disabled, FreeDR adds 16,384 tokens to each reply's output limit.
• Reasoning is never stored or shown. FreeDR keeps it only in memory, so that it can be sent back to the provider during tool calls: OpenAI encrypted reasoning, Anthropic signed thinking blocks and DeepSeek-style reasoning_content.
• More thinking is slower and uses more tokens.

API key storage

• Each AI provider has its own write-only key slot. The four AI slots sit alongside the threat-intel provider keys. The FreeDR service encrypts the keys with Windows DPAPI under the service account and stores them in the control database.
• The console shows only whether a key is saved. It never shows the key value. You can keep, replace or remove a key with Save changes. Switching provider never copies a key to another provider.
• FreeDR rejects enabling AI for a provider that has no saved key.
• You can also enter a key in the setup wizard, or import one from the optional setup configuration file (JSON).
• FreeDR never shows provider error bodies, URLs or keys in its error messages. It shows only fixed messages such as external service HTTP 429; retryable=true; automatic retry disabled.

Settings

Setting (UI label)	Default	Meaning
Enable AI investigation	off	Master switch for investigations and assists.
AI provider / Model ID / Thinking	CheaperInference / deepseek-v4.1-flash / medium	See above.
Daily token limit	20,000,000	Allowed range is 400,000 to 20,000,000 tokens per UTC day.
Let AI contain threats it finds	off	Lets the investigator run its four containment actions. When off, its response plan is only a draft.
Enable public web research	off	Gives the investigator web_search and read_public_page. Also enables manual research from a case, which needs no AI key.

Token accounting and budgets

FreeDR counts tokens, not money. It does not calculate or show prices. Your provider's own billing applies.

Daily cap. All AI work counts against one cap per UTC day: investigations, explanations and drafts. The default and maximum is 20,000,000 tokens.

Investigations.

• Up to 3 investigations run at once, and at most one per case.
• When an investigation starts, it reserves an equal share of the remaining daily headroom across the free run slots. It needs at least 50,000 tokens to start. Unused tokens go back to the budget when the run ends.
• Investigations always leave 96,000 tokens free (2 × 48,000), so explanations and drafts still work while investigations hold reservations.
• A run is limited to 30 rounds and 100 tool observations. One request can be at most 400,000 tokens.
• Every tool result shows the run's remaining tokens. The model is told to finish while one more request still fits.

Assists. Explanations and drafts are single requests:

• Up to 2 assists wait on the provider at once.
• One assist can reserve at most 48,000 tokens.
• Output caps: alert explanation 8,192 tokens, exclusion draft 2,048, Hunt and rule drafts 4,096. Thinking adds 16,384 to each cap.

Billing rules.

• When the provider reports usage, that number replaces the reservation.
• Some failures happen before the provider does billable work: the request is too large, the client cannot start, there is no connection, or the provider returns HTTP 401, 403 or 429. These release their reservation.
• If billing is unknown (timeout, interrupted response, restart), FreeDR keeps the tokens counted. For an investigation this is at most one request's worth.
• A reply cut off at the output limit is billed and discarded.

What the run shows. Each run shows tokens used (input and output) and tokens reserved.

Budget errors.

• If the budget is used up, requests fail with ai_daily_budget_exhausted.
• If running work holds the remaining headroom, requests fail with ai_busy.

Retries.

• The HTTP client never retries on its own.
• The investigator retries transient provider failures with backoff: 2, 5, 15, 30, then 60 seconds, repeating until the run stops or its budget no longer fits. Transient failures are timeouts, lost connections, HTTP 429 and server errors. The run detail shows, for example, "provider HTTP 503; retrying in 15 s".
• A reply it cannot parse is retried once.

What is sent to the provider

What the model receives.

• The deterministic case summary, which is the same summary you see on the case.
• Tool results it asks for.
• A fixed system prompt that includes the evidence manual.

Redaction. FreeDR redacts locally supplied content before every request:

• These fields and columns are replaced with [redacted]: password, secret, api_key, authorization, access_token, refresh_token, host_id, username. Account-name columns (user_name, actor_user_name) are not redacted, so the model sees names such as NT AUTHORITY\SYSTEM.
• SIDs (S-1-…, in any field) are handled by kind:
  • Well-known SIDs are kept: SYSTEM S-1-5-18, LocalService S-1-5-19, NetworkService S-1-5-20, built-in groups such as Administrators S-1-5-32-544, service SIDs S-1-5-80-…, Everyone S-1-1-0, and integrity labels S-1-16-…. These name no person or machine, so the model can see who ran something.
  • Machine or domain accounts (S-1-5-21-…) with a well-known RID below 1000 keep only the RID: S-1-5-21-[redacted]-500 is the built-in Administrator.
  • User SIDs (S-1-5-21-… with RID 1000 or more), Microsoft Entra ID SIDs (S-1-12-…) and any other SID become [redacted SID].
  • Only a value that is a whole SID is changed; a SID inside a longer text (for example an HKU\S-1-5-21-…\ registry path) is sent as is.
• Any string that contains credential markers becomes [redacted credential-bearing text]. The markers include password=, --password, api_key=, bearer, token=, secret=, ghp_ and github_pat_.
• In C:\Users\<name>\… paths, the user name becomes <user>.

Redaction is pattern-based

Other text in the evidence still goes to the provider: command lines, file paths, registry values, script blocks and host details. The provider's own terms and data retention apply. FreeDR does not claim zero data retention.

Web research. When web research is enabled, search terms go to https://search.parallel.ai/mcp, an anonymous search service, and page reads go to the public URL itself. See Web research.

Where AI results appear

Place	What you see
Case → AI tab	Runs, live tool history and token usage, the report, response actions the AI ran, and the draft.
Alert detail	Explain this alert.
False-positive / Tune this rule dialog	Ask AI to draft.
Hunt	Describe what to find and Draft as rule.
Overview, case view and rule detail	Pending exclusions filed by the AI, with Approve / Reject.
Case → Response tab	AI-submitted jobs, under the purpose AI investigator (run <id>): ….

API endpoints

All AI endpoints are on the local API. They need the local session, and every POST also needs the CSRF token.

Method and path	Purpose
GET /api/v1/ai/status	Shows whether a provider is configured for assists.
GET /api/v1/cases/{case_id}/ai	Lists a case's runs (limit 1–100, default 20, cursor).
POST /api/v1/cases/{case_id}/ai	Starts an investigation (question up to 2,048 bytes; optional previous_run_id for a follow-up).
GET /api/v1/cases/{case_id}/ai/{run_id}	Gets one run.
POST /api/v1/cases/{case_id}/ai/{run_id}/cancel	Stops a run.
POST /api/v1/cases/{case_id}/research	Runs a manual web search (query, up to 512 bytes) or page read (url, up to 2,048 bytes).
POST /api/v1/detections/alerts/{alert_id}/ai-explanation	Explains an alert. Returns the cached explanation unless regenerate is set.
POST /api/v1/detections/alerts/{alert_id}/ai-exclusion	Drafts an exclusion's field choice.
POST /api/v1/hunt/ai-draft	Drafts Hunt SQL (target: hunt) or a custom rule (target: rule).

Error codes:

Code	Meaning
invalid_ai_request	The request is malformed.
ai_busy	Slots or budget are held by running work.
ai_disabled	AI is switched off.
ai_unavailable	The service or provider cannot be reached.
ai_daily_budget_exhausted	The daily token cap is used up.
ai_case_or_session_changed	The case or the investigation session changed.
ai_output_rejected	The model's reply did not match the required shape and was discarded.