HEARTH THOR Collective
Open · MIT · Community curated

Keep the
hearth burning.

HEARTH is an open library of threat hunting hypotheses — drafted by hunters, refined by AI, and organized around the PEAK framework. Browse what works. Paste a CTI link, get a hunt. Contribute one back.

—Hypotheses
—Contributors
3PEAK categories
—ATT&CK techniques mapped
01 — PEAK Framework

Three ways
to hunt.

Every hunt in HEARTH is classified into one of three categories from the PEAK threat hunting framework. They're not levels — they're different muscles. Use the right one for the job in front of you.

01 · Flames —

Hypothesis
driven

A specific theory about adversary behavior, and the evidence you'd expect to find. Classic threat hunting — write it down, then go look. "An adversary uses DLL side-loading via a signed app to maintain persistence."

—→
02 · Embers —

Baseline
& explore

Understand "normal" before you can spot the anomaly. Inventory parent-child process trees. Map the long tail of user agents. Find what you didn't know you had.

—→
03 · Alchemy —

Model-
assisted

Algorithmic and ML-powered approaches. Clustering, statistical baselines, outlier detection — let the data surface threats you wouldn't think to look for.

—→
02 — From CTI to hunt

Paste the link.
Get a hunt.

Drop a threat intel report URL into the pipeline. HEARTH fetches the article, summarizes the TTPs, validates technique IDs against MITRE ATT&CK, checks the library for duplicates, and drafts a complete hypothesis you can review, edit, and merge.

hearth pipeline · run-2049 · 00:00:14 elapsed

$ hearth ingest https://mandiant.com/resources/blog/apt-living-off-the-land

→ resolving · brotli · readability-lxml fallback

✓ content extracted · 4,218 words · 12 IOCs

→ analyzing with claude-sonnet-5 · drafting hypothesis…

tactic : TA0003 Persistence

technique : T1547.001 Registry Run Keys / Startup

confidence: 0.92 · MITRE validated

✓ hypothesis drafted · 312 words · PEAK = Flames

→ checking 162 hunts for similarity…

! 2 weak matches found · jaccard 0.31, 0.27

✓ opened PR #0231 · awaiting Keeper review

draft · H-0231-flames.md
Hunt H-0231 · Flames · drafted by Claude

Persistence via signed-binary side-loading from a non-standard %APPDATA% path.

Hypothesis. An adversary establishes persistence by dropping a malicious DLL adjacent to a signed Microsoft binary in a user-writable directory, then registering the binary as a Run key — bypassing application allow-lists that trust the signer rather than the path.

TacticPersistence
TechniqueT1547.001
Sub-techT1574.002
03 — Actor coverage

Got a target?
Check the coverage.

Name a threat actor. HEARTH shows which hunts cover them, where coverage runs thin, and which of their techniques nobody has hunted yet.

hearth · /actors
live · drawn from public/context-graph-data.json

Match HEARTH to your threat model.

Type a name or alias. We'll surface every hunt that covers that group's tradecraft, score the coverage tactic by tactic, and call out the techniques where HEARTH is still blind — with a one-click path to fill them in.

04 — Hunt library

Browse, fork,
and burn it down.

Every hypothesis is a versioned markdown file — readable, forkable, runnable. Start with the featured hunt of the week, or filter the full library by tactic, data source, or PEAK category.

Hunt of the week · 19 May 2026

OAuth token replay across tenant boundaries.

by L. Proehl · reviewed by S. Marrone · 14 forks

A stolen refresh token is replayed from a new ASN within minutes of issuance, against a tenant the original user has never accessed. The hypothesis: legitimate token use clusters tightly by ASN and geography within the first 24 hours — replay shows up as a sharp deviation.

Flames T1550.001 Entra ID Cloud
Open full hunt →
Detection query · KQL
// Token replay outside expected ASN cluster
SigninLogs
| where ResultType == "0"
| where AuthenticationProtocol == "oAuth2"
| summarize
    asns = make_set(NetworkLocationDetails),
    first = min(TimeGenerated)
  by UserId, SessionId
| where array_length(asns) > 1
| where first > ago(24h)
| project UserId, SessionId, asns, first
Required data
SigninLogs AADNonInteractive CloudAppEvents
Hypothesis
PEAK
Technique
Tactic
Contributor
Loading library…
05 — In your AI assistant

Ask what
to hunt.

The HEARTH MCP server puts the whole library inside Claude, Cursor, VS Code and any other MCP client. Ask in plain language by keyword, threat actor, or the logs you actually have. It's read-only, sends no telemetry, and treats community hunt text as untrusted data.

install · one command

# Claude Code

$ claude mcp add hearth -- npx -y @thorcollective/hearth-mcp

# Claude Desktop, Cursor, VS Code: see the setup guide

 

> you

What can I hunt for Volt Typhoon with firewall and appliance logs?

→ hunts_for_actor · Volt Typhoon (G1017)

data sources : network, appliance

runnable now : 9 hunts · 83 more need another source

coverage : 55 / 81 ATT&CK techniques

four read-only tools
search_huntsRanked search by keyword, tactic or data source.
get_huntA full hunt: hypothesis, notes, queries, sources.
hunts_for_actorHunts, coverage and gaps for a threat actor.
hunts_for_datasourcesWhat your team can run with the logs it has.
06 — Keepers & contributors

It takes
a collective.

HEARTH is built by working hunters. The Keepers maintain the framework and review every PR — and — contributors and counting have submitted hunts from the trenches of their own SOCs.

Top contributors · 90 daysHunts
Keeper · co-founder
Lauren Proehl

"A hypothesis written down is worth ten in someone's head. HEARTH is the place we write them down — together."

@jotunvillur
Keeper · co-founder
Sydney Marrone

"Threat intel without action is decoration. HEARTH turns CTI into something you can actually run on Monday."

@letswastetime