Repository navigation
AiSOC v23.2.0
Added
-
Per-IOC threat-intel feed lookup, wired into scoring and investigation.
The confidence scorer has always recognised feed-sourced TI hits —misp,
otx,taxii,kev— but nothing populated those slots: the feeds landed
in the threat-intel stores while the scorer readfused.enrichments, which
only the enrichment service writes. A newThreatIntelClientin
services/enrichmentasksservices/threatintel's existing per-IOC search
on every IP, domain, URL and hash, and a match surfaces under the matching
feed's own source name, so a feed hit now moves the alert's confidence and
appears in investigations. Two tenant-propagation gaps are closed with it:
fusion'sAlertEnrichercarries the alert's tenant on each bulk item, and
the investigator'senrich_iocforwards the run's tenant bound by the
registry closure — never an LLM-visible argument. The lookup URL is built
positionally (url.URL+url.Values), confining the attacker-influenced
IOC value to the encoded query string. -
Sophos Central connector.
sophos_centralreads Sophos Central alerts
— endpoint and server threats, exploit and ransomware protection, intrusion
prevention — throughGET /siem/v1/alerts, the feed Sophos's own SIEM
integration reads (#356). There is no region to configure: the client
credential is exchanged at Sophos ID,whoaminames the tenant and its data
region host, and a partner or organization credential reads the one tenant
itstenant_idnames, found through the tenant lookup. A data region host
outsidesophos.comis refused rather than sent the token, because it
comes from a response and the egress guard only vets configured values.
The window is read to the end throughread_token_pages, which gains a
more_field: Sophos returnsnext_cursoron its last page too and says
whether more follow withhas_more, so ending on the token alone would
re-request the last page up to the page cap. The SIEM API refuses a
from_dateolder than 24 hours, which the scheduler's outage backfill asks
for, so a longer window is shortened to fit and logged as
sophos_central.window_shortened. Severity keeps the vendor's LOW, MEDIUM
and HIGH and never emitscritical; events are filed as OCSF 2001 Security
Findings, so the generic-mapping count stays at 42. Setup, partner tenants
and troubleshooting are on the new
Sophos Central page. -
A released CHANGELOG section is now frozen to its tag.
scripts/check_changelog_frozen.pycompares the newest five released
sections byte-for-byte against theCHANGELOG.mdcommitted at each
section's own tag, with a unified diff on mismatch, an explicit refusal
when a tag cannot be resolved, and a vacuous-pass refusal when nothing
could be compared. It exists because the same drift shipped three times in
two days: a fix PR writes its entry under an empty[Unreleased], the
release promotes that section, and when the PR merges after the tag the
three-way merge anchors the entry by context — inside the released
section, claiming a fix the tagged artifact does not contain (Datadog into
[23.0.0], then Cortex and Trend into[23.1.0]). Run against the
pre-fix tree the gate exits 1 naming[23.1.0]and showing both drifted
entries; this change also moves those entries back where they belong,
below. Writing the gate found a fourth, older instance: a post-tag credit
line appended to the released[22.1.0]section. That section is restored
to its tag's bytes and the credit lives in.github/CREDITS.md, the
surface built to stay current. -
A saved connector can be edited from the console. Configure on a
connector card was disabled ("planned for v1.1") and its handler answered
"Delete + re-add", while the GitHub, GitLab, Falco, Tines, Torq and Entra
pages already told operators to rotate credentials through the AiSOC
connector edit screen. Re-adding a connector costs its id, its checkpoint
(the first poll re-reads the whole lookback), its event counters and, for
push connectors, the ingest token the sender is configured with. The new
EditConnectorModalbuilds its form from the catalog schema, prefills the
stored settings, starts every credential blank and sends only what changed:
a typed credential replaces its own key, a blank one is kept, a cleared
optional setting is removed. To make that safe,
PATCH /api/v1/connectors/{id}now mergesauth_configand
connector_configinto the stored row one top-level key at a time — a
supplied key replaces,nullremoves, an omitted key survives — under a
row lock, so the scheduler's checkpoint write cannot be overwritten by a
concurrent edit. Before this, both blobs were replaced wholesale: re-typing
one secret erased the others, a change to one setting erased the
checkpointandfilter_rulesbeside it, and{"auth_config": {}}erased
every credential. An API client that relied on the replacement to drop a
key now sends that key asnull. A blanknameis refused with 422
instead of being stored. -
Shodan, URLScan and IPInfo enrichment is real. The enrichment service
readSHODAN_API_KEY,URLSCAN_API_KEYandIPINFO_API_KEYfrom its
config and did nothing with any of them — an operator setting a key got a
silent no-op whilePortInfosat documented as "detected by Shodan" with
nothing filling it. Three clients now fill that gap: Shodan host lookups
(open ports, tags, hostnames), URLScan verdicts and IPInfo geolocation/ASN,
each wired into the enricher chain with transport-level tests, compose and
env documentation. One correction to the contributed code: Go's
*url.Errorstringifies the full URL it failed on, so a key in the query
string reached the logs through any wrapped transport error — IPInfo's
token moves to the vendor-supportedAuthorization: Bearerheader, and
Shodan (whose contract requires?key=in the URL) redacts the query
before wrapping, both pinned by tests. Contributed by @43n51c
(#1831, carried in
#1835).
Fixed
-
Cortex XDR and Cortex XSIAM read 100 incidents per poll and lost the
rest. Both connectors askedPOST /public_api/v1/incidents/get_incidents
forsearch_from: 0, search_to: 100and stopped there. The endpoint pages by
offset — a request returns the incidents fromsearch_fromup to
search_to, at most 100 of them, with no link or token to the next page —
and Palo Alto's own integration reads page n assearch_from = n * 100.
A poll window opens where the last one closed, so a window holding more than
100 incidents lost the rest for good. Both now move the offsets on until a
page comes back short, through a new sharedread_offset_pagesin
app/connectors/_paging.py, which keeps what was read when a later page
fails and announces the 25-page bound when it binds, as the link and token
readers do. The read is now ordered oldest-first (creation_time asc): an
incident created during the read joins the end of the list instead of
shifting every offset beneath the pages still to be read. A refused opening
request behaves as before — Cortex XDR raises so the scheduler records a
failed poll, Cortex XSIAM logscortex_xsiam.fetch_failedand returns
nothing. The paging gate recognises the new helper, andMAX_UNCLASSIFIED
falls from 52 to 50. -
Trend Vision One read one page of each feed per poll and lost the rest.
TrendVisionOneConnector.fetch_alertsread the first response of
GET /v3.0/workbench/alertsand ofGET /v3.0/oat/detectionsand ignored
nextLink, the complete URL of the next page that both feeds return. Trend's
own clients —get_itemsin its API cookbook and thepytmv1SDK — keep
requesting that URL until a response no longer carries one. A poll window
opens where the last one closed, so a window holding more workbench alerts or
OAT detections than one page lost the rest for good. Both feeds now read
throughread_linked_pages, which requests the link verbatim with the same
token, keeps what was read when a later page fails and announces the 25-page
bound when it binds. A refused or failed opening request on either feed is
logged astrend_vision_one.feed_failed/trend_vision_one.feed_exception
with the feed named, and the other feed is still read; a refused OAT request
was previously dropped without a log line.MAX_UNCLASSIFIEDfalls from 50
to 49. -
Settings, Case and Responder deep links now survive navigation.
/settings?tab=deploymentopened Profile because the section was a
hardcodeduseState('profile'); the case workspace read?tab=only as a
mount-time initializer, so tab clicks, query changes and browser history
drifted apart; and/responderredirected every installed-PWA shortcut to
Triage, eating the?tab=queue|cases|oncallthe manifest publishes. A
shareduseUrlTabhook validates tab values against each surface's
allowlist in both directions, dedups history entries, and preserves other
query parameters and the fragment. Contributed by @knowlet
(#1817). -
The agents service never received
METRICS_TOKEN, so:15002/metrics
answered 401 on every default deployment. The endpoint gates on the
token exactly as the api service's does, but only the api block in
docker-compose.ymlcarried the passthrough — the Prometheus triage
counters were unreachable out of the box. Reported by @pitimon with the
confirming override (#1822,
fixed in #1823). -
A connector's
error_countnow means "consecutive failed polls since
the last success".record_poll_failureincremented it and
record_poll_successnever reset it, so the fleet panel rendered a
lifetime total beside a freshlast_syncas if the connector were failing
right now — the reported case showed a static "19 errors" for days on a
connector that ingested 237k events during the observation window.
Success writes 0; failure still increments; both writers are pinned by
tests. Reported by @pitimon
(#1821, fixed in
#1824). -
The Wazuh indexer reads follow
search_afterto the end of the
window. Both fetch paths sent onesize: 1000search and stopped,
under a docstring calling the bound deliberate — but the cut was silent,
and because the next poll's window opens where this one closed, the
overflow was never re-requested by anything. The shared_searchnow
follows OpenSearch's documented continuation with a_doctiebreaker so
equal timestamps cannot skip or repeat rows, keeps partial reads when a
later page fails, stops rather than looping when a full page carries no
sort values, and announces the 25-page bound when it binds.
MAX_UNCLASSIFIEDfalls from 49 to 48. Reported by @pitimon from code
inspection, stated as such
(#1820, fixed in
#1825). -
The console discarded usable refresh tokens when access expired.
Protected JSON, streaming and download requests now renew the session through
the existing refresh endpoint, share concurrent renewal, and retry a rejected
access token once. Temporary refresh failures preserve the session; rejected
credentials still require sign-in. Late responses cannot restore a logged-out
session or replay an action under a different login. Desktop and responder
guards recognize recoverable sessions, and the responder no longer lets a
stale expiry cache overrule a renewed JWT. The Explain stream also receives
its missing proxy rule to the agents service. Token lifetimes and server-side
revocation checks are unchanged. -
Copilot chat is grounded in caller-scoped security data instead of
answering from prose alone. The chat transports bind three read-only
tools — alerts, one record by UUID, Fusion entity risk — that forward the
caller's own authorization and tenant to the core API, which remains the
authority: no LLM argument or page context can choose a URL, credential or
tenant;kindis enum-bound,record_idround-tripsuuid.UUIDwith a
response-echo check, entity paths are percent-encoded, redirects are
disabled and reads bounded. The text cache is disabled when tools are
bound (a cached prose answer must not impersonate a live read), a
truncated tool loop no longer publishes its last raw lookup as a
conclusion, and the Fusion risk endpoints answer 503 instead of
fabricating an empty queue. Contributed by @knowlet
(#1819, carried in
#1828). -
The frozen-changelog gate no longer reds the release PR it exists to
protect. Promoting[Unreleased]to[X.Y.Z]creates a released
section whose tag can only exist after the release PR merges, which the
two-day-old gate reported as UNVERIFIABLE and fatal — every future
release would have failed its own required check. The newest released
section, when its version equals the working tree'sVERSIONfile and
its tag does not resolve, is now reported as an explicit PENDING state
instead of a failure; the exception is bounded on every side (an untagged
section anywhere else stays fatal, a pending version whose tag already
exists is compared like any other, a newest section not matching
VERSIONstays fatal) and each bound is pinned by its own test. -
The Agentmetry inbox template preserves the context an investigation
needs. Selected account provenance, supporting-event references,
recorder coverage and MCP schema observations now survive normalisation
underfinding.evidence.agentmetry, with reported account claims kept
separate fromactor.userand full account emails and opaque account/user
IDs excluded. Field paths, the activity/finding split, the severity ladder
and the OCSF 1.1.0 declaration are unchanged
(#1826, from discussion
#1577).
Security
-
The mobile app's
node-forgefloor rises to 1.4.3. All four advisories
against 1.4.2 (Ed25519 small-order acceptance, two TLS-client crash paths,
ReDoS in header trimming) name 1.4.3 as the first patched version; the
existing override floor moved from>=1.4.1and exactly one package moved
in the re-resolved lock, at both import sites. Closes Dependabot alerts
#232–#235 (#1829). -
The session live check validates its fixture token's shape. The script
that exercises session refresh against a live stack now refuses a fixture
credential that is not structurally a JWT before sending it anywhere,
closing CodeQL alert #1129
(#1830).