Turns one-shot external recon into a continuous monitoring program. Covers the scheduled re-scan-and-diff loop (baseline snapshot -> interval sleep -> re-scan -> asset/finding delta -> threshold-gated webhook alert), the scan-to-scan diff engine (new/removed/changed assets by a tracked-attribute table, new/resolved findings by a stable cross-scan fingerprint), adversary CTI / chatter monitoring across six public feeds (ransomwatch, ransomware.live, HackerNews Algolia search, Reddit security-subreddit RSS, GitHub Gist code-search, public Telegram channel scraping) with a source-kind-aware severity engine (leak-site/forum/telegram/paste tiers, CRITICAL through INFO), literal/glob/regex watchlist pattern matching, full-corpus capture with retroactive rescan on new watchlist entries, infrastructure-tracking-over-time discipline (certificate-transparency, passive-DNS, port/service, and typosquat re-enumeration cadence, and what a genuine 'perimeter drift' event looks like in the diff output), a five-state finding-
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Turns one-shot external recon into a continuous monitoring program. Covers the scheduled re-scan-and-diff loop (baseline snapshot -> interval sleep -> re-scan -> asset/finding delta -> threshold-gated webhook alert), the scan-to-scan diff engine (new/removed/changed assets by a tracked-attribute table, new/resolved findings by a stable cross-scan fingerprint), adversary CTI / chatter monitoring across six public feeds (ransomwatch, ransomware.live, HackerNews Algolia search, Reddit security-subreddit RSS, GitHub Gist code-search, public Telegram channel scraping) with a source-kind-aware severity engine (leak-site/forum/telegram/paste tiers, CRITICAL through INFO), literal/glob/regex watchlist pattern matching, full-corpus capture with retroactive rescan on new watchlist entries, infrastructure-tracking-over-time discipline (certificate-transparency, passive-DNS, port/service, and typosquat re-enumeration cadence, and what a genuine 'perimeter drift' event looks like in the diff output), a five-state finding-lifecycle state machine (open/triaged/risk_accepted/resolved/false_positive) with per-severity SLA and fingerprint-based cross-scan dedup and auto-resolve/reopen rules, the alert-fatigue trap where a lifecycle-unaware rule re-fires on an already-accepted finding, a durable retry/backoff alert-outbox pattern ('queued is not delivered'), and copy-paste bash-cron plus PowerShell-Scheduled-Task recipes for a re-scan+diff loop with the Slack-compatible webhook payload shape. Passive OSINT and analysis only -- no new active-intrusion technique. Use when setting up ongoing monitoring for a retainer or MSSP engagement, tuning alert thresholds to avoid fatigue, triaging a finding's lifecycle status, investigating adversary chatter about a brand, building a 'what changed on the perimeter since last week' report, or deciding whether a persisting finding should re-alert.
version
1
sources
asm_reference_impl, public_research
triggers
["continuous monitoring","continuous exposure monitoring","retainer monitoring","MSSP monitoring","scheduled rescan","scheduled scan","scan diff","diff scans","monitor a target","monitor continuously","re-scan and diff","delta alert","alert on delta","drift detection","attack surface drift","perimeter drift","what changed since last scan","ransomware leak site monitoring","leak site monitoring","adversary chatter","dark web monitoring","brand mention monitoring","IAB monitoring","paste site monitoring","telegram brand monitoring","CTI feed","threat intel feed","chatter watchlist","retroactive rescan","infrastructure tracking over time","certificate transparency monitoring","CT log monitoring","[Truncated]"]
Companion skills: osint-methodology (the 5-stage pipeline this skill
loops — see its §7.2 "ongoing weekly diff" profile, which this skill fills in with concrete
mechanics), offensive-osint (§29 Threat Intel & IOCs — this skill
deepens that section's advisory/IOC-feed directory with the continuous adversary-chatter
watch loop and CTI-feed cadence it explicitly lacks; use §29 for indicator enrichment and
vulnerability-prioritization data sources, this skill for the standing collection loop),
org-attack-surface (the org-first discovery this skill's re-scans
re-run on a schedule), exposure-risk-quantification (reads
this skill's finding-lifecycle suppression state to compute risk_trend and the FAIR score — see
its risk-score model). This skill answers a different question than all four: not "what does the
target expose right now," but "is what the target exposes changing, and should anyone be told."
0. When to Use / When NOT
Use this skill when:
Standing up ongoing monitoring for a retainer, MSSP, or bug-bounty program instead of a one-shot
engagement — the client wants to know about new exposure, not to re-read yesterday's report.
Deciding how often to re-run which recon stage (daily vs. weekly vs. monthly) without either
wasting API quota / detection budget on cheap-to-skip stages or missing real drift.
Building or tuning adversary-chatter monitoring (ransomware leak sites, forum/paste mentions,
Telegram brand mentions) for a target's brand/domain.
Deciding whether a finding that keeps showing up in every re-scan should keep alerting, or has
already been triaged/accepted and should go quiet.
Debugging "the webhook never fired" or "the webhook fired twice" — alert delivery reliability,
dedup, and backoff behavior.
Writing a "what changed on the perimeter since last week" deliverable.
Do NOT use this skill when:
You need the one-shot discovery methodology itself — that's osint-methodology (5-stage
pipeline) and offensive-osint (the per-technique arsenal). This skill assumes discovery already
happened at least once and is about the run.
second and every subsequent
You need a new active-intrusion technique. This skill is a scheduling, diffing, and alerting
layer over recon that is already authorized and already running — see §5.
The target's authorization isn't established, or a recurring cadence hasn't been explicitly
agreed — see §1's monitoring-specific authorization note.
1. Authorization & Legal Posture
Same base posture as the companion skills: intended for assets the operator owns or has written
authorization to assess — see osint-methodology §1 for the full soft-scope-check script.
The monitoring-specific nuance: a point-in-time engagement authorization does not automatically
cover an indefinite recurring cadence. Before standing up a monitor/chatter watch daemon or a
dashboard-scheduled job against a target, confirm the authorization window explicitly covers the
monitoring period (retainer end date, RoE renewal terms), not just "the engagement." A scheduled job
that outlives its authorization is a standing violation, not a one-off mistake — it fires every
interval until someone notices and tears it down.
Always-on guardrail specific to this skill: a monitoring loop must never silently escalate its
own authorization tier. The monitor CLI command's flag surface only exposes --only/--exclude
module filters — it has no path to enable --validate (the credential/token-submission tier), so a
scheduled asm-cli monitor loop cannot accidentally re-arm that tier. The dashboard job
scheduler is different: a job's flags field accepts the same flags a manual scan does, so an
operator can configure a recurring job with --validate/--validate-creds set. Treat that as a
standing-authorization decision that needs its own explicit sign-off, not a monitoring-cadence
decision — see §5.
2. Confidence Levels
Same three-tier scale as the companion skills — TENTATIVE / FIRM / CONFIRMED — answers "how sure
am I this is real." This skill adds one axis that is easy to conflate with confidence and must not
be:
Confidence is not lifecycle status. A chatter hit can be HIGH severity and still TENTATIVE
confidence (adversary-controlled content is always TENTATIVE by design — see §7.6). A finding can be
CONFIRMED confidence and still be risk_accepted in the lifecycle (§9) — the client acknowledged a
real exposure and chose to keep it. Never let a severity or confidence label imply anything about
whether a human has already triaged the finding, and never let a lifecycle status imply anything
about whether the underlying evidence was independently verified. They are orthogonal axes recorded
in different places (the Finding's confidence field vs. finding_lifecycle.status).
3. Output Format
Three schemas this skill's outputs use. All timestamps UTC ISO-8601.
Finding — same schema as the companion skills (see osint-methodology §3); chatter-sourced
findings always carry confidence: tentative (§7.6).
Delta event — what a monitoring cycle reports:
DeltaEvent:
scan_a_id / scan_b_id: older / newer scan
scan_a_time / scan_b_time
kind: new_asset | removed_asset | changed_asset | new_finding | resolved_finding
asset_or_finding_key: the asset key, or the finding's fingerprint::asset_key::title
type_or_category: asset type, or finding category
severity: info|low|medium|high|critical (findings only)
changes: {attr: [old, new]} (changed_asset only)
alert_fired: bool — did this event cross the configured threshold
Lifecycle record — cross-scan identity for one finding at one target (see §9):
LifecycleRecord:
target: lowercased registrable domain
fingerprint: 16-hex sha1, stable across scans (§9.2)
status: open | triaged | risk_accepted | resolved | false_positive
first_seen_ever: UTC, earliest scan_started_at this fingerprint appeared in
last_present_scan/at: most recent scan that reproduced this fingerprint
age_scans: count of scans this fingerprint has appeared in
sla_due: UTC, computed from severity at first insert (§9.4)
triaged_by/at, resolved_by/at, notes
4. Source Hygiene & Citations
Same discipline as the companion skills: URL + UTC timestamp + SHA-256 + tool version + run_id per
artifact. One monitoring-specific addition: the chatter corpus tables (ransomware_corpus) persist
every observed row regardless of whether it currently matches a watchlist, precisely so retroactive
analysis (§7.4) has real historical data to search rather than needing to re-fetch a since-rotated
feed. Treat that corpus as an evidence store in its own right — dedup on the natural key
(source/group/victim), never delete rows to "clean up," and cite corpus_id alongside the hit id
when a finding traces back to a retroactive match.
5. Do NOT
Do not treat a scheduled job's "queued" status as "delivered" — see §10.1. Confirm delivery by
querying the outbox, not by the enqueue call succeeding.
Do not re-alert on a finding whose lifecycle status is resolved, false_positive, or
risk_accepted just because a re-scan rediscovered it — see §9.6 for the exact mechanics and which
alerting subsystem actually protects against this (only one of the two does, natively).
Do not let a recurring monitoring job silently carry --validate/--validate-creds unless
that specific standing authorization was explicitly confirmed — see §1.
Do not open a chatter-sourced URL (leak site, paste, Telegram post) directly in your working
browser. Adversary-controlled content is TENTATIVE by definition (§2, §7.6) — open only in an
isolated browser/VM/Tails, exactly as osint-methodology §6.1 (OpSec) requires generally.
Do not invent an SLA day-count, backoff schedule, or poll interval. This skill's numbers (§9.4,
§10.2) are transcribed from the shipped defaults; if an operator wants different numbers, say so
explicitly rather than presenting a made-up default as authoritative.
Do not widen a findings_watchlist rule's severity_min or drop its category/module
filters to "catch more" without first reading §9.6 and §10.5 — an unscoped rule on this subsystem
re-fires every re-scan on a persisting issue, unlike the fingerprint-diff alerting in §6.
Do not introduce a new active-intrusion technique to "monitor better." Every module a
monitoring loop re-runs is a technique that was already authorized for the one-shot engagement;
this skill only adds cadence, diffing, and alerting around it.
6. The Monitoring Loop Model
6.1 The five-step loop
The canonical pattern (monitor.py):
On startup, run a scan immediately — establishes a baseline if none exists.
Sleepinterval seconds.
Run a new scan.
Diff the latest two scans for the target (§6.3 / §7).
If the delta crosses the alert threshold, POST a webhook (§10). Otherwise log and continue.
Go to 2.
A first run against a fresh target has no prior scan to diff against — the loop logs "first scan —
no prior baseline to diff against; skipping alert" and simply establishes the baseline. Nothing
alerts on run #1 by design.
6.2 Two implementations
A mature ASM implementation ships two ways to run this loop — pick based on scope:
CLI foreground daemon
Dashboard job scheduler
Entry point
asm-cli monitor <target> ...
monitor_scheduler.scheduler_daemon_loop
Scope
one target per process
many jobs, persisted in MonitorJobStore
Visibility
own stdout only
dispatched via ScanManager.launch() — appears in the live dashboard scan list
Poll/dispatch cadence
the loop's own --interval
polls for due jobs every 30s (_POLL_SECONDS)
Per-run safety cap
none — interval only
7200s (2h) _SCAN_MAX_SECONDS; a wedged scan is cancelled and the loop moves on (already rescheduled)
Disable
Ctrl-C
ASM_DISABLE_MONITOR_SCHEDULER=1
--validate reachable?
No — the CLI's flag surface is --only/--exclude only
Yes — job flags accepts the same flags a manual scan does; see §1
Both funnel through the same diff-and-alert logic (monitor._maybe_alert) — the scheduler literally
imports and calls it, so §6.5's alert-worthy rule is identical either way.
6.3 What "diff" means
The diff engine (reporting/diff.py) compares two scan.db files and produces:
New / removed assets — pure set difference on asset key.
Changed assets — for keys present in both scans, a per-type tracked-attribute table decides
what counts as a meaningful change (deliberately narrow — noise like response time or timestamp
casing would otherwise drown the real signal):
A note on existence-only types: since an asset's key typically embeds its value (e.g.
ip:1.2.3.4), a subdomain re-pointing to a new IP does not surface as a "changed" event — it
surfaces as a paired removed_asset (old IP) + new_asset (new IP), because the key itself
changed. Read a re-point out of the new/removed pairing, not out of changed_assets.
New / resolved findings — keyed by the finding's persisted fingerprint (v6+ scan.db; older
DBs fall back to a synthetic asset_key::title key). A finding present in scan B but not scan A is
"new"; present in A but not B is "resolved." Fingerprint mechanics — including why a resolved
finding reappearing is not silently treated as fine — are in §9.
6.4 Cadence guidance
Not every stage is worth re-running on the same schedule. Passive re-enum is cheap; active probing
costs quota, time, and detection budget (osint-methodology §6.4 back-off ladder still applies to
every active re-run).
Pairs with whatever cadence you picked for a full re-scan
Same as the full-scan cadence
6 — --validate tier
credential/token-submission modules
Highest authorization tier
Never on an unattended schedule without a human re-confirming scope each run — see §1
Rule of thumb: if the client's driver is "tell me about new subdomains, certs, buckets, and
ransomware chatter," a daily stage-1/2-only loop plus the chatter daemon (§7) covers it cheaply.
If the driver is "tell me about new exposed misconfigurations," you need stage 4, and weekly is the
realistic floor once asset count is non-trivial.
6.5 What's alert-worthy
monitor._maybe_alert fires when either:
A new finding's severity rank is at or above the configured --threshold (default medium), or
A new asset's type is in the hot-asset-type set: credential, typosquat_domain, bucket,
repo — these fire regardless of severity threshold, because a brand-new credential leak,
lookalike domain, exposed bucket, or newly-discovered repo is always worth a look even before
triage assigns it a severity.
Nothing else in the diff — resolved findings, changed assets, new low-severity findings outside the
hot-asset types — triggers a webhook on its own. It's still in the diff output and the Markdown
report (§6.3), just not pushed.
7. CTI / Adversary Chatter Monitoring
The chatter/ subsystem is the standing adversary-side watch: six public sources polled on their own
cadence, matched against an operator watchlist, classified into severity, and (above threshold)
pushed to a webhook.
https://www.reddit.com/r/<sub>/.rss across cybersecurity, netsec, InfoSecNews, blueteamsec, Pwned, AskNetsec, sysadmin
600s
none
paste:gist
paste
https://api.github.com/search/code (filtered to gist.github.com results)
900s (15 min)
GITHUB_TOKEN
telegram
telegram
https://t.me/s/<channel> public web preview
1800s (30 min)
none, but needs channels configured under market_intel.telegram.channels (seed with asm-cli breach discover --kind telegram)
The two leak-site sources (ransomwatch, ransomware_live) each poll their full feed on every
cycle and persist every row to a local ransomware_corpus table — not just watchlist matches.
That full-corpus capture is what makes §7.4's retroactive rescan possible.
hackernews and reddit always inject a set of generic seed keywords (data breach, ransomware,
0day, supply chain attack, vulnerability, CVE for HN; a broader breach/leak/exploit list for
Reddit) alongside the operator's watchlist, so the chatter dashboard shows ambient threat-landscape
chatter even before a watchlist exists — those hits are tagged relevance: ambient and are excluded
from client-facing promotion (§7.6).
7.2 Watchlist pattern types
Every source routes matching through a shared compiler (chatter/matching.py). Three pattern types,
all case-insensitive:
Type
Semantics
Default
literal
substring match
yes
glob
fnmatch wildcards (*, ?, [seq])
regex
full Python regex, search-mode
A glob or regex pattern that fails to compile falls back to literal match with a logged warning — one
bad pattern can never wedge the whole collection cycle. Add a pattern:
asm-cli chatter watchlist add "acme*.com" --kind domain --pattern-type glob \
--priority high --notes "primary + all TLD variants"
asm-cli chatter watchlist add "acme*.com" --kind domain --pattern-type glob `
--priority high --notes "primary + all TLD variants"
--kind is one of domain | brand | keyword | email — it feeds the severity engine below (a
domain-kind match ranks higher than a brand/keyword match at the same source).
7.3 The severity engine
chatter/classifier.py::classify() — never raises, always returns a (severity, category, classification, reason) tuple. The rubric is source-kind-aware, not a flat rule:
Source kind
Signal
Severity
leak_site
domain-kind match + a named ransomware group recognized in the text
CRITICAL
leak_site
any other match on a leak-site listing
HIGH
forum
IAB-style phrase (see below) + domain match + a price/currency marker in the excerpt
HIGH
forum
IAB phrase + domain match, no price
MEDIUM
forum
domain match, no IAB phrase
MEDIUM
forum
brand/keyword match
LOW
telegram
secret-regex hit in the post text
HIGH
telegram
domain match + IAB phrase
CRITICAL
telegram
domain match, no IAB phrase
HIGH
telegram
email match
HIGH
telegram
brand/keyword match
MEDIUM
telegram
any other match
MEDIUM (floor — adversary channels never rank below this)
paste / other
secret-regex hit in the excerpt
HIGH
paste / other
domain match + IAB phrase
HIGH
paste / other
domain match, no IAB phrase
MEDIUM
paste / other
brand/keyword match
LOW
paste / other
email match
MEDIUM
paste / other
no rule fired
INFO
Note the telegram tier floors one step above the equivalent paste-site tier for every pattern kind —
an adversary-run channel is treated as inherently higher-signal than an ambient paste dump.
Recognized ransomware groups (partial — checked case-insensitively against the post text):
lockbit, akira, cl0p/clop, ransomhub, play, blacksuit, medusa, qilin, rhysida,
8base, cactus, inc ransom, hunters international, darkvault, embargo, dragonforce,
kasseika, plus legacy names still seen on mirrors: revil, conti, alphv/blackcat,
blackbyte, royal, vice society, bianlian, snatch.
A hit's severity can also be promoted — never demoted — by a source-supplied severity_hint
(HackerNews/Reddit infer high/medium/low/info from title keywords like "leak"/"breach"/"CVE"/
"advisory" before the classifier even runs) when the hint outranks the classifier's own verdict.
7.4 Retroactive rescan on new watchlist entries
Adding a watchlist entry (chatter watchlist add, default rescan=True) doesn't just start matching
future chatter — it immediately searches the full historical ransomware_corpus for matches
already sitting in the corpus:
Literal patterns use an indexed LIKE search (fast, unbounded).
Glob/regex patterns pull the most recent 10,000 corpus rows and filter in Python via the shared
matcher (bounded so a pathological pattern can't stall the CLI/daemon).
Matches found this way are inserted as normal chatter_hits rows but are pre-marked alerted=1 —
the daemon will not webhook-spam the operator with 18-month-old news the moment a new watchlist
entry is added. This is what solves "the client was hit by this group over a year ago and we only
just started watching for them" — the historical hit shows up in the dashboard and the hit-detail
view, silently, without an alert storm.
7.5 Full-corpus capture
Both leak-site sources persist every row they see on every poll to ransomware_corpus, dedup'd on a
natural key, regardless of whether anything currently on the watchlist matches. This is the
precondition for §7.4 and for any later trend-tracking query ("how many of our client's peers in this
sector have appeared on a leak site in the last 90 days") — without full-corpus capture, a watchlist
added today can only ever see chatter from today forward.
7.6 Promotion to scan.db
chatter/promote.py::promote_hit() — a one-way copy of a chatter hit into a scan's asset+finding
tables, mirroring the dorking-results promotion pattern:
Asset:AssetType.WAYBACK (the existing "externally-harvested URL" type), tagged
sources=["chatter:<source>"].
Finding:Category.INFO_DISCLOSURE, severity from the classifier, confidence always
TENTATIVE — chatter is adversary-controlled content by definition, and the description text
explicitly instructs: "Open the URL ONLY in an isolated browser / VM / Tails. Validate
authenticity before escalating."
Only fires for hits at/above min_finding_severity (default medium) — low/info hits stay
asset-only, visible in the dashboard, not promoted as a finding.
Ambient hits are hard-blocked from promotion. A hit whose relevance is ambient (matched only
a generic seed keyword, no watchlist pattern behind it — §7.1) returns PromotionResult(ok=False, ...) unconditionally. There is no client-specific signal behind an ambient hit, so it must never
reach the client-facing deliverable or scan.db findings — only a genuine watchlist match can.
7.7 Running it
asm-cli chatter watch --webhook https://hooks.slack.com/services/XXX/YYY/ZZZ \
--threshold high --tick 30
asm-cli chatter watch --webhook https://hooks.slack.com/services/XXX/YYY/ZZZ `
--threshold high --tick 30
--threshold (default high) is the severity floor for the webhook, independent of §6.5's
monitor threshold — they're two separate daemons with two separate knobs (§10.5). --tick is how
often the daemon checks whether any source's per-source cadence (the table in §7.1) has elapsed; it
does not itself poll every tick.
8. Infrastructure Tracking Over Time
This methodology assumes no push/streaming CT-log feed — every signal in this section is poll +
diff: re-run the relevant module on the loop's cadence (§6.4) and read the delta out of §6.3's
asset diff. What follows is which module produces which tracked-attribute signal, and what it means
when it fires.
A subdomain's IP changing shows as a paired removed+new ip asset (§6.3's existence-only-type note), not a "changed" event
New open port / service change
ports (stage 3 — Shodan InternetDB passive, or --probe active native scan)
New port asset (tracked ip/port/transport); a fingerprint/product change on an existing port is a changed_assets entry (tracked ip/port/product on service)
New typosquat / lookalike domain
typosquat (stage 2 — dnstwist permutations + DNS resolution of candidates)
New typosquat_domain asset — always fires the webhook per §6.5's hot-asset-type rule, regardless of severity threshold; a mail-capable lookalike is escalated to HIGH by the module itself
The single highest-signal "perimeter drift" event class this platform ships is a new
typosquat_domain, credential, bucket, or repo asset — all four are in monitor.py's
hot-asset-type override (§6.5), meaning they alert regardless of the configured severity threshold.
If a client asks "what's the one thing I should never have muted," it's these four asset types.
8.1 A worked weekly recipe
Either let the loop do it automatically (§6, §11) or run it by hand:
# Last week's scan id vs. this week's, most-recent-first in scans/
asm-cli diff acme_com_20260730 acme_com_20260806 \
--scans-root scans --out reports/weekly-diff.md --json reports/weekly-diff.json
The Markdown report (reporting/diff.py::render_markdown) sections, in order: Summary counts → 🆕
NEW findings (sorted highest severity first) → ✅ RESOLVED findings → 🆕 NEW assets (grouped by type)
→ ➖ REMOVED assets (grouped by type) → 🔄 CHANGED assets (with the old→new value for every changed
attr). Each section caps at 50 items with a …N more ellipsis so a noisy diff doesn't blow out the
report.
9. Finding Lifecycle & False-Positive Discipline
The fleet/ subsystem is what lets a finding have a state that persists across scans, instead of
every re-scan treating every finding as new.
9.1 The five states
open → triaged → {risk_accepted | resolved | false_positive}
↑____________________________| (resolved can REOPEN to open — see §9.3)
open, triaged, risk_accepted, resolved, false_positive — enforced enum
(_LIFECYCLE_STATUSES). Two of the five are suppressed — excluded from dashboard counters, the
FAIR risk grade, and attack-path chains: false_positive and resolved.
risk_accepted is deliberately not suppressed. It's a real, acknowledged exposure the client
chose to keep — it must still count toward the risk grade and still appear in attack-path reasoning.
This is the single most common lifecycle mistake to correct: "the client accepted the risk" does not
mean "this finding no longer matters for scoring," it means "the client made an informed decision
about a finding that is still real."
normalized_title = lowercased, all digit runs replaced with #, whitespace collapsed. The
digit-normalization is deliberate: a title like "Open port 8080 exposed" and "Open port 9090 exposed"
both normalize to the same string (open port # exposed), so the same underlying misconfiguration
type on the same asset stays one lifecycle identity even as incidental digits in the title drift
between scans. The \x1f (ASCII unit separator) between fields avoids cross-field collisions from
any field that happens to contain a more common delimiter.
Two related-but-different keys exist — do not conflate them:
fleet_findings is keyed (scan_id, fingerprint) — one row per scan, the per-scan snapshot.
finding_lifecycle is keyed (target, fingerprint) — one row per target, the cross-scan
identity that actually carries the state machine.
9.3 Reconciliation — auto-resolve and reopen
Runs automatically at the end of every scan (fleet.promote.promote_scan_into_fleet →
FleetStore.reconcile_lifecycle), given the full set of fingerprints present in the just-completed
scan:
New fingerprint (never seen for this target) → INSERT status='open', age_scans=1, sla_due
computed from severity (§9.4).
Fingerprint reappears in a newer scan → age_scans += 1, last_present_scan/at bumped. If its
status was resolved, it is reopened to open — a resolved issue coming back is a real
regression and must not be silently treated as still-fine.
Fingerprint absent from the new scan, and its status was open or triaged, and it was last
present in a strictly earlier scan (same-scan idempotence guard) → auto-resolved
(resolved_by='auto').
risk_accepted and false_positive fingerprints are untouched by reconciliation either way — they
don't auto-resolve when absent (nothing to resolve, the operator already made a call) and they don't
silently flip back to open on reappearance the way resolved does; they stay exactly as the
operator left them unless a human changes the status.
9.4 SLA
sla_due is computed once, at first insert, from the finding's severity:
Severity
SLA (days)
critical
7
high
14
medium
30
low
90
info
none
fleet/sla.py::sla_state() classifies a lifecycle row against now: overdue (active status, past
sla_due), due_soon (within 3 days of due), on_track, or none (closed status, or an
info-severity finding with no SLA at all). Only open/triaged rows can be overdue or
due_soon — a risk_accepted/resolved/false_positive row is never flagged against the SLA clock
regardless of how old it is.
9.5 Suppression feeds risk scoring
fleet.store.suppressed_fingerprints(target) — the single source of the suppression rule — is
consumed by reporting/risk_score.py and reporting/attack_paths.py. A false_positive or
resolved fingerprint is invisible to both; a risk_accepted one is visible to both (§9.1). See
exposure-risk-quantification for how that feeds the FAIR score and risk_trend.
9.6 The alert-fatigue trap — which subsystem actually protects you
This is the answer to the recurring question "why did I get alerted again on a finding we already
marked risk-accepted / false-positive":
Two separate alerting subsystems exist, and only one of them is lifecycle-aware:
Subsystem
Dedup key
Lifecycle-aware?
monitor / monitor_scheduler diff alerting (§6)
the finding's fingerprint — a re-scan that reproduces the same fingerprint is not a "new" finding, full stop
Yes, structurally — same fingerprint can never appear in new_findings twice
findings_watchlist rule matching
(rule_id, finding_id) where finding_id is the per-scan database row id, freshly generated by every scan
No — _matches_rule() checks severity/category/module/target-pattern/title-regex/KEV/EPSS only; it never reads finding_lifecycle. A fresh scan produces a fresh finding_id for the same underlying fingerprint, so the (rule_id, finding_id) dedup key does not recognize it as a repeat
The practical consequence: a risk_accepted finding that keeps reappearing in every re-scan will
not re-trigger a monitor webhook (its fingerprint isn't "new"), but a broad
findings_watchlist rule with a low severity_min and no category/module narrowing will
re-fire on it every single time the daemon's cursor sees a scan that reproduces it — because from
that rule's point of view, it's a fresh finding row it hasn't alerted on before.
Recommendations, in priority order:
Prefer monitor's fingerprint-diff alerting for "tell me about new things" — it's the only one
with lifecycle-equivalent protection built in.
If you need findings_watchlist rules (they're the tool for "alert on any finding matching X,
in any scan," which diff alerting can't express), narrow them hard — specific category +
module, or require_kev/epss_min — rather than a blanket severity_min=medium catch-all.
If a findings_watchlist rule keeps re-firing on a triaged issue, the fix is upstream: mark the
fingerprint's lifecycle status in fleet (§9.1) and, if the rule is too broad to respect that
automatically, tighten the rule rather than muting the whole webhook.
10. Alert Delivery Reliability
10.1 "Queued is not delivered"
Every alert sender in this platform (monitor, chatter daemon — which literally imports and reuses
monitor.py's _post_webhook, so its outbox rows show up tagged kind="monitor", not
kind="chatter" — and findings_watchlist, tagged kind="watchlist") enqueues into a durable
outbox (alerts/store.py) instead of POSTing inline. The sender's own success/notified bookkeeping
reflects enqueue succeeding, not delivery. findings_watchlist._post_alert says so explicitly in
its docstring — "returns (True,'') on enqueue (queued = success for the caller's notified
bookkeeping)" — and monitor._post_webhook behaves identically (it enqueues and returns without
confirming delivery). To know whether a webhook
actually reached its destination, query the outbox itself — never infer it from the caller's log
line.
Delivery attempts back off 120s → 240s → 480s → 960s → 1920s → 3840s → 7680s (the min(…, 21600)
6h cap is a theoretical ceiling the default max_attempts=8 never reaches — 7680s is the last backoff
actually scheduled). After 8 attempts (max_attempts, default), the row's status flips to dead — a literal dead-letter
state, visible in the outbox for manual retry, not silently dropped.
enqueue() skips inserting a new row if a row with the same dedup_key is still pending —
duplicate-suppression only against not-yet-delivered duplicates, not against history (a genuinely
repeated event after the first one delivered will enqueue again, correctly).
10.4 Delivery worker and retention
alert_worker_loop runs every 30s (alert_worker_interval), delivers due rows, and — roughly
once per day — purges delivered rows older than 14 days (alert_retention_days) so the outbox
doesn't grow unbounded on an unattended box. Manual controls: retry(id) and retry_all_failed()
reset attempts to 0 and status back to pending (available from the dashboard's outbox page).
10.5 Threshold tuning per subsystem — avoid alert fatigue
Subsystem
Threshold knob
Default
What it gates
monitor (CLI)
--threshold
medium
New-finding severity floor; hot-asset types (§6.5) always fire regardless
monitor_scheduler (dashboard job)
job's alert_threshold field
medium
Same _maybe_alert() logic, per job
chatter watch
--threshold
high
Chatter-hit severity floor for the webhook
findings_watchlist rule
severity_min field, plus category/module/target_pattern/title_regex/require_kev/epss_min
high
Per-rule; narrow with the extra filters rather than relying on severity alone (§9.6)
Start conservative (high-severity-only, plus the always-on hot-asset types) and widen only after
confirming the loop is quiet at the tighter setting — it's far easier to notice "I'm not getting
enough signal" and loosen a threshold than to diagnose alert fatigue after a client starts ignoring
the channel.
11. Scheduler Recipes
A mature ASM implementation ships three native alternatives to a bare cron loop — prefer these where they fit
before reaching for external scheduling:
Dashboard job scheduler — multi-target, appears in the live dashboard, has the 2h per-run
safety cap (§6.2); configure through the dashboard's monitoring page.
asm-cli chatter watch --webhook <url> --threshold high — the adversary-chatter daemon
(§7.7).
A bare cron/Task-Scheduler loop is still the right call for: ad-hoc periodic reports without wanting
a standing foreground process, wrapping scan+diff+report inside an existing ops pipeline, or a shared
box where a long-running unsupervised Python daemon isn't operationally convenient.
11.1 bash cron — weekly re-scan + diff
# crontab -e (or drop a file under /etc/cron.d/)
# Every Monday 03:00 — re-scan, diff against last week's scan, write both Markdown + JSON.
0 3 * * 1 cd /opt/asm-cli && \
OLD_SCAN=$(ls -1t scans | grep '^acme_com_' | sed -n 1p) && \ # newest EXISTING dir = last run; captured before the new scan below is created
NEW_SCAN="acme_com_$(date +%Y%m%d)" && \
asm-cli scan acme.com --scan-id "$NEW_SCAN" --scans-root scans \
>> logs/weekly.log 2>&1 && \
asm-cli diff "$OLD_SCAN" "$NEW_SCAN" --scans-root scans \
--out "reports/diff_$(date +%F).md" --json "reports/diff_$(date +%F).json" \
>> logs/weekly.log 2>&1
Adjust --only/--exclude on the scan line per §6.4's cadence table if you want a cheap daily
stage-1/2-only cron alongside a heavier weekly full re-scan (two separate cron lines, two separate
scan-id prefixes).
11.2 PowerShell Scheduled Task — same recipe on Windows
# Register once (run as the service account that owns the asm-cli install):
$action = New-ScheduledTaskAction -Execute "powershell.exe" -Argument `
'-NoProfile -ExecutionPolicy Bypass -File C:\asm-cli\weekly-diff.ps1'
$trigger = New-ScheduledTaskTrigger -Weekly -DaysOfWeek Monday -At 3am
Register-ScheduledTask -TaskName "AsmWeeklyDiff" -Action $action -Trigger $trigger
All three alert senders (monitor, chatter, findings_watchlist) emit a Slack-compatible
payload — a text field (Slack Incoming Webhooks render it directly) plus a structured envelope any
generic JSON-accepting webhook handler can read. The monitor shape:
The chatter shape swaps the summary block for a single-hit envelope (severity, category,
classification, target, source, group_name, title, url, hit_id, first_seen_at,
posted_at); the findings_watchlist shape nests alert.rule and alert.finding (including
evidence, so a KEV/EPSS-gated rule's payload carries the CTI context that triggered it) instead of a
scan-pair summary. All three keep the plain-text text field first so a handler that ignores
everything else still renders something legible.
12. Skill Self-Test
Drop these into a fresh session to verify the skill loads and routes correctly.
"Set up ongoing monitoring for acme.com — new subdomains, certs, and ransomware chatter, cheaply." → §6.4 (stage 1/2 daily), §7.7.
"What's the difference between asm-cli monitor and the dashboard job scheduler?" → §6.2.
"A subdomain's IP address changed between two scans. Does the diff show that as 'changed'?" → §6.3 (no — paired removed+new ip asset, existence-only types don't track attrs).
"What real endpoint does the ransomware.live source hit, and how often?" → §7.1.
"I just added a watchlist entry for a brand we were breached under 18 months ago. Will this spam our webhook with old news?" → §7.4 (no — retroactive hits pre-marked alerted).
"A Telegram post mentions our domain plus 'RDP access' — what severity?" → §7.3 (CRITICAL — domain + IAB phrase on telegram).
"Should an ambient HackerNews hit about 'ransomware' with no watchlist match get promoted into the client's scan.db?" → §7.6 (no — ambient hits are hard-blocked from promotion).
"A finding we marked risk_accepted a month ago still shows up in every weekly re-scan. Is it counted in the FAIR risk grade?" → §9.1, §9.5 (yes — risk_accepted is not suppressed).
Negative test:"We marked a finding risk_accepted. Should it re-alert the next time the monitor loop diffs a re-scan that still contains it?" → §9.6 (no, for monitor's fingerprint-diff alerting — same fingerprint isn't 'new'. But flag the caveat: a broad findings_watchlist rule with no category/module narrowing WILL re-fire, because it dedups on the per-scan finding_id, not the fingerprint.)
"A resolved finding just reappeared in this week's scan. What happens to its lifecycle status?" → §9.3 (reopened to open — a regression, not silently ignored).
"Our webhook sender logged 'alert fired' but the client says they never got a Slack message. How do I check what actually happened?" → §10.1 (query the outbox — queued ≠ delivered).
"What's the retry schedule before an alert is marked dead?" → §10.2 (60s base, ×2 each attempt, capped 6h, dead after 8 attempts).
"Give me a weekly cron job that re-scans and diffs a target, PowerShell version too." → §11.1, §11.2.
"Can I set a monitoring job on the dashboard to include --validate so it re-runs credential testing every week automatically?" → §1, §6.2 (technically reachable via job flags — but treat it as a standing-authorization decision, not a cadence decision, and confirm sign-off first).
"Show me the webhook payload shape for a monitor delta alert." → §11.3.
13. Changelog
v1.0 (2026-08-06) — initial release. The five-step monitoring loop and its two native
implementations — CLI foreground daemon vs. dashboard job scheduler (§6); the scan-to-scan diff
engine's tracked-attribute table and alert-worthy delta rule (§6.3–§6.5); a stage-by-stage
re-scan cadence guide (§6.4); adversary-chatter monitoring across all six shipped public sources
with exact endpoints/cadences, the source-kind-aware severity engine, retroactive rescan-on-new-
watchlist-entry, full-corpus capture, and the ambient-hit promotion gate (§7); infrastructure-
tracking-over-time mapped to the concrete module + tracked-attribute signal behind each drift
event (§8); the five-state finding-lifecycle machine, fingerprint mechanics (including the
digit-normalization rationale), auto-resolve/reopen reconciliation, SLA table, and — the
highest-value original finding in this skill — the precise reason findings_watchlist rules are
NOT lifecycle-aware while monitor's fingerprint-diff alerting structurally is, and what that
means for avoiding re-alerts on triaged findings (§9); the durable alert-outbox pattern including
the "queued is not delivered" distinction, backoff/dead-letter numbers, dedup key, and retention
(§10); and copy-paste bash-cron + PowerShell-Scheduled-Task recipes plus the Slack-compatible
webhook payload shape for all three alert senders (§11). Every number, endpoint, cadence, and
rule in this skill is transcribed from a production ASM implementation's shipped monitoring,
chatter, scan-diff, fleet, findings-watchlist, and alert-outbox modules — not invented.
