merge: testing → main (reconcile 2-week divergence)

This commit is contained in:
2026-04-28 18:36:00 -04:00
parent 499836c9e4
commit 862e4dbb31
1235 changed files with 160255 additions and 7996 deletions

View File

@@ -0,0 +1,313 @@
"""Similarity-graph primitives for the connected-components clusterer.
Each function takes two :class:`Observation` projections and returns a
similarity score in ``[0.0, 1.0]``. The connected-components impl
(landing in subsequent commits) decides how to combine these into a
single edge weight, applies a threshold, and runs union-find.
**Time-agnostic.** Edges MUST NOT depend on observation timestamps.
Fixture 7 (``slow_burn``) proves recency-decay clustering fragments
multi-month APT campaigns; the production graph cannot silently expire
old edges. Timestamps are still useful for *audit* (the ``first_seen``
on the resulting identity row) but never for *similarity*.
**Weight tiers** (from `development/IDENTITY_RESOLUTION.md`):
* High — JA3 / HASSH / payload-hash / C2-callback exact match. Stable
signals an attacker can't cheaply rotate. A single high-tier match
supports identity strongly.
* Medium — command-sequence Jaccard, bucketed by UKC phase. Tooling
habits leak through command order; phase-bucketing avoids comparing
a Discovery cmd-list to an Exploitation one.
* Low — credential-attempt-set Jaccard. Defeated alone by fixture 1
(``shared_wordlist``) where two campaigns share rockyou but diverge
on infra.
* Very low — ASN match. Defeated alone by fixture 2 (``vpn_hopping``)
where one identity rotates across many ASNs.
The functions are pure (no DB, no I/O); the worker maps observations
into :class:`Observation` once per tick and feeds these into the
graph builder.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Mapping, Optional
# ─── Observation projection ─────────────────────────────────────────────────
@dataclass(frozen=True)
class Observation:
"""Minimal projection of a per-IP attacker observation.
Built once per ``Attacker`` row by the worker (or per
``SyntheticAttacker`` in tests via :func:`from_synthetic`).
Keeping the projection tight isolates the graph code from schema
drift on either side.
All set-typed fields are :class:`frozenset` so they hash and so
callers don't accidentally mutate them mid-pass.
"""
observation_id: str
"""Stable ID — for production, the ``Attacker.uuid``; for tests,
the ``SyntheticAttacker.attacker_id``."""
ja3: Optional[str] = None
hassh: Optional[str] = None
asn: Optional[int] = None
payload_hashes: frozenset[str] = field(default_factory=frozenset)
c2_endpoints: frozenset[str] = field(default_factory=frozenset)
credentials: frozenset[tuple[str, str]] = field(default_factory=frozenset)
commands_by_phase: Mapping[str, tuple[str, ...]] = field(default_factory=dict)
"""``UKCPhase.value`` → ordered command sequence observed in that
phase. Empty dict when no command-bearing sessions were seen."""
# ─── Edge functions ─────────────────────────────────────────────────────────
def _fingerprints_fully_disagree(a: Observation, b: Observation) -> bool:
"""True iff every comparable fingerprint slot disagrees.
"Comparable" = both sides have a non-null value for that slot.
Used as a soft-veto on shared C2 / payload signals: when two
observations have distinct stable TLS + SSH stacks, sharing a C2
endpoint is a *campaign*-level signal (cooperating operators,
distinct identities) — not an identity-level one. Fixture 5
(``multi_operator``) is the canonical demonstration.
Returns ``False`` when no fingerprint slot is comparable (any-null
cases) — without evidence of disagreement we don't veto. Also
``False`` when at least one slot agrees.
"""
ja3_comparable = a.ja3 is not None and b.ja3 is not None
hassh_comparable = a.hassh is not None and b.hassh is not None
if not (ja3_comparable or hassh_comparable):
return False
if ja3_comparable and a.ja3 == b.ja3:
return False
if hassh_comparable and a.hassh == b.hassh:
return False
if ja3_comparable and hassh_comparable:
return a.ja3 != b.ja3 and a.hassh != b.hassh
return True # exactly one slot is comparable, and it disagrees
def high_weight_edge(a: Observation, b: Observation) -> float:
"""JA3 / HASSH / payload-hash / C2-endpoint exact match.
Returns ``1.0`` if any of the four exact-match signals agrees
(non-null on both sides), ``0.0`` otherwise. Single-signal high-tier
agreement is by design enough to support identity — these are the
signals the design doc calls out as "stable signals an attacker
can't cheaply rotate."
**Fingerprint-disagreement veto.** Payload and C2 are infra signals
that two cooperating operators (different identities) can share.
JA3 + HASSH are tooling signals that differ when the operators are
actually different humans with different tool stacks. So when the
available fingerprint slots fully disagree, we drop the
payload/C2 contribution to zero — preventing a campaign-level
co-op signal from fusing two distinct identities. Fixture 5
(``multi_operator``) is the canonical demonstration: shared
stage-1 payload + shared C2, distinct JA3/HASSH per operator —
must stay two identities. JA3 / HASSH agreement still returns
``1.0`` directly, since by definition no veto applies when
something agrees.
JA4 will join this tier as a sibling of JA3 once the prober emits
it (``ATTACKER_FINGERPRINTED`` already carries a JA4 slot in
``AttackerIdentity``); the function shape doesn't change.
"""
if a.ja3 is not None and a.ja3 == b.ja3:
return 1.0
if a.hassh is not None and a.hassh == b.hassh:
return 1.0
if _fingerprints_fully_disagree(a, b):
# Stable-tool disagreement vetoes shared-infra signals.
return 0.0
if a.payload_hashes and b.payload_hashes and (a.payload_hashes & b.payload_hashes):
return 1.0
if a.c2_endpoints and b.c2_endpoints and (a.c2_endpoints & b.c2_endpoints):
return 1.0
return 0.0
def medium_weight_edge(a: Observation, b: Observation) -> float:
"""Phase-bucketed command-sequence Jaccard.
For each UKC phase observed on both sides, computes the Jaccard
similarity of the command sets (multisets collapsed to sets — the
*order* signal is reserved for a future feature, this commit is
the scaffolding). Returns the **maximum** Jaccard across shared
phases, so a single strong phase match isn't averaged away by a
different phase where the actors diverge.
Phase-bucketing matters: comparing a Discovery cmd-list to an
Exploitation one is meaningless. Both actors had to be in the
same phase for the comparison to count.
Returns ``0.0`` when no phase is observed on both sides.
"""
shared_phases = set(a.commands_by_phase) & set(b.commands_by_phase)
if not shared_phases:
return 0.0
best = 0.0
for phase in shared_phases:
sa = set(a.commands_by_phase[phase])
sb = set(b.commands_by_phase[phase])
if not sa and not sb:
continue
union = sa | sb
if not union:
continue
j = len(sa & sb) / len(union)
if j > best:
best = j
return best
def low_weight_edge(a: Observation, b: Observation) -> float:
"""Credential-attempt-set Jaccard.
Returns the Jaccard of ``(username, password)`` tuples. Two campaigns
burning the same wordlist will score high here — fixture 1 proves
this signal is dangerous in isolation. The connected-components
impl combines this with other signals; alone it must not push a
pair over threshold.
Returns ``0.0`` when either side attempted no credentials, or when
the union is empty.
"""
if not a.credentials or not b.credentials:
return 0.0
union = a.credentials | b.credentials
if not union:
return 0.0
return len(a.credentials & b.credentials) / len(union)
def very_low_weight_edge(a: Observation, b: Observation) -> float:
"""ASN equality.
Returns ``1.0`` iff both observations have a non-null ASN and they
match. Fixture 2 (``vpn_hopping``) proves ASN-only clustering is
a failure mode — one identity legitimately rotates across many
ASNs. The combination logic in the connected-components impl
weights this so that ASN agreement alone never crosses threshold.
"""
if a.asn is None or b.asn is None:
return 0.0
return 1.0 if a.asn == b.asn else 0.0
# ─── Combined weight ────────────────────────────────────────────────────────
#: Tier multipliers applied to the per-tier edge scores when combining
#: into a single weight. Tuned so that:
#:
#: * High-tier agreement alone (1.0) crosses the 1.0 threshold.
#: * Medium-tier alone (max 1.0) yields 0.6 — below threshold.
#: * Low-tier alone (max 1.0) yields 0.2 — defeats fixture 1's
#: credential-overlap-only failure mode.
#: * Very-low alone (max 1.0) yields 0.05 — defeats fixture 2's
#: ASN-rotation failure mode.
#:
#: The ratio between tiers matters more than the absolute values: a
#: tier should never combine its way past threshold without help from
#: a stronger one.
TIER_WEIGHTS = {
"high": 1.0,
"medium": 0.6,
"low": 0.2,
"very_low": 0.05,
}
#: Threshold a combined edge weight must meet to survive into the
#: similarity graph. The connected-components impl drops anything
#: under this before running union-find.
EDGE_THRESHOLD = 1.0
def combined_edge_weight(a: Observation, b: Observation) -> float:
"""Sum of all four tier scores, weighted by :data:`TIER_WEIGHTS`.
Each per-tier function returns a score in ``[0, 1]``; the
weighted sum lets stronger tiers dominate without letting weaker
ones combine their way past threshold.
The connected-components clusterer compares this against
:data:`EDGE_THRESHOLD` to decide whether to draw an edge. Pure /
time-agnostic — fixture 7 forbids recency-decay weighting.
Commits 57 land each tier in the call site:
* Commit 5 (this commit): high + medium.
* Commit 6: + phase-handoff (a separate edge family, not a tier).
* Commit 7: + low + very_low.
Until commit 7 lands, the low / very_low contributions stay zero
by virtue of the underlying functions returning ``0.0`` whenever
their inputs are missing. The combination is forward-compatible.
"""
return (
TIER_WEIGHTS["high"] * high_weight_edge(a, b)
+ TIER_WEIGHTS["medium"] * medium_weight_edge(a, b)
+ TIER_WEIGHTS["low"] * low_weight_edge(a, b)
+ TIER_WEIGHTS["very_low"] * very_low_weight_edge(a, b)
)
# ─── Adapter for the synthetic-corpus tests ─────────────────────────────────
def from_synthetic(att) -> Observation: # type: ignore[no-untyped-def]
"""Build an :class:`Observation` from a ``SyntheticAttacker``.
Lives here so test code doesn't import the factory shape into the
production module — the adapter is a documented integration point.
Imported lazily by callers; the production worker uses a parallel
adapter from :class:`Attacker` rows once that lands.
"""
payload_hashes: set[str] = set()
c2_endpoints: set[str] = set()
credentials: set[tuple[str, str]] = set()
commands_by_phase: dict[str, list[str]] = {}
for s in att.sessions:
if s.payload_hash:
payload_hashes.add(s.payload_hash)
if s.c2_callback:
c2_endpoints.add(s.c2_callback)
for cred in s.credentials_tried:
credentials.add(tuple(cred))
if s.commands:
commands_by_phase.setdefault(s.phase.value, []).extend(s.commands)
return Observation(
observation_id=att.attacker_id,
ja3=att.ja3,
hassh=att.hassh,
asn=att.asn,
payload_hashes=frozenset(payload_hashes),
c2_endpoints=frozenset(c2_endpoints),
credentials=frozenset(credentials),
commands_by_phase={k: tuple(v) for k, v in commands_by_phase.items()},
)
__all__ = [
"Observation",
"high_weight_edge",
"medium_weight_edge",
"low_weight_edge",
"very_low_weight_edge",
"combined_edge_weight",
"from_synthetic",
"EDGE_THRESHOLD",
"TIER_WEIGHTS",
]