Cold Open: Morning Lockout and the 40,000-Attempt Trail
The incident surfaced not through an overnight pager escalation, but when a colleague opened a laptop, attempted an interactive session on our Windows Server 2019 Remote Desktop Protocol (RDP) jump host, and encountered an active account lockout. Forensic inspection of the Security log established that the host had absorbed roughly 40,000 failed logon attempts (Event ID 4625) overnight while our detection and alerting pipeline remained silent.
Although no silicon failed or thermal limit was breached, the overnight flood imposed measurable physical resource contention across the host. Each incoming TCP SYN on port 3389 traversed a stale CIDR block persisting in our perimeter allowlist, completed the three-way handshake, and routed CredSSP, NTLM, or Kerberos authentication negotiation into the Local Security Authority Subsystem Service (LSASS.exe). Within LSASS, every invalid credential set triggered account resolution, cryptographic hash or ticket verification, Local Security Policy lockout-counter updates, and structured audit-record serialization into the Security event ring buffer. Sustained at high concurrency, those operations diverted CPU cycles from session-broker scheduling, increased storage I/O as .evtx blocks flushed and rotated on disk, and incremented the bad-password counter until the target account crossed its lockout threshold—causing subsequent valid credentials from an authorized engineer to fail with a lockout status code.
This failure mode stems from two interacting design flaws: perimeter boundary drift and alert under-tuning under lean staffing. A stale allowlist entry exposes an administrative choke point to external scanners, while suppressing individual Event ID 4625 alerts to spare a two-person team from notification fatigue—without deploying rate-based burst thresholds or automated perimeter or host enforcement—renders high-volume credential stuffing indistinguishable from baseline quietude until an operational lockout exposes the breach.
Perimeter Drift: How a Stale Allowlist Left RDP Exposed
A stale CIDR block on a perimeter firewall or cloud security group fails silently: the rule remains active and continues forwarding packets that now map to foreign topology. When an ISP reclaims a netblock, a contractor departs, a branch office migrates carriers, or a VPN egress pool shifts addresses, the legacy /24 or /28 routinely persists in production access control lists to avoid touching live rules outside a maintenance window. Because automated scanners continuously sweep IPv4 space on TCP port 3389, the moment a forgotten prefix is reassigned to a recycled cloud pool or an adversary-controlled range, every inbound probe matching that CIDR completes a TCP three-way handshake directly into the jump host's network stack.
Once the handshake completes and Remote Desktop Protocol (RDP) negotiation begins, work shifts from stateless packet matching to stateful host execution. Inbound SYN-ACK completions and RDP transport packets pass through tcpip.sys into Remote Desktop Services (TermService) and CredSSP, routing authentication payloads into LSASS.exe. Each invalid logon forces LSASS.exe to resolve the principal, execute NTLM challenge-response or Kerberos/CredSSP validation involving cryptographic hash computations and security context checks, increment the per-account bad-password counter in the SAM or domain security state, and serialize a Windows Security Event ID 4625 record into the event log channel. Across 40,000 failed attempts overnight—roughly 55 attempts per minute over twelve hours—those repeated context switches, LSASS.exe authentication loops, and .evtx file flushes sustain elevated CPU scheduling overhead and disk write amplification while ratcheting bad-password counters toward the account lockout threshold.
Where Allowlist Drift Actually Opens the Door
- Reassigned ISP and cloud CIDR blocks: Carriers and cloud providers routinely recycle unallocated or abandoned IP ranges that once belonged to a partner site or site-to-site gateway. External TCP SYN sweeps to port 3389 match the stale allow entry and receive SYN-ACK responses from the jump host, exposing an unauthenticated RDP listener directly to the public internet without an intervening VPN or bastion hop.
-
Oversized masks masking a single host: Administrators frequently provision a
/24or/16allow rule for a single NAT gateway IP, leaving 253 or 65,533 adjacent addresses inside the permitted boundary. Compromising any neighboring host or virtual machine within that prefix gives an attacker direct, uninspected RDP reachability to the jump host. - Split-horizon and stale jump-chain rules: Rules originally written for an internal management segment or retired bastion host sometimes get duplicated into edge policies or left enabled after a routing cutover. Asymmetric or misrouted paths then carry external traffic into the jump host listener, turning an internal administrative path into an external ingress vector.
- Host firewall drift after policy edits: Even when an upstream firewall or security group blocks a stale CIDR, an overly broad local Windows Defender Firewall rule permitting TCP 3389—or a temporary diagnostic rule left in place—admits traffic whenever a secondary network interface, route, or NAT translation reaches the host.
Burst Thresholds and Alerting That Won't Bury a Two-Person Team
Paging on every individual Event ID 4625 fails immediately on a two-person team: routine typing errors, stale cached credentials, and mistyped usernames generate continuous isolated 4625 events, causing operators to mute or ignore single-event alerts within days. Effective detection evaluates rate-based burst windows over rolling intervals—typically 1-minute or 5-minute buckets—incrementing per-key counters (by source IP or target account) and firing an alert only when sustained volume crosses a defined threshold, followed by cooldown suppression so a single scan does not flood the on-call queue.
| Scope | Typical Burst Threshold | Why This Shape Works for a Lean Team |
| Per Source IP | 15–25 failed logons in 1 minute (or 40+ in 5 minutes) | Automated scanners and credential-stuffing workers hammering TCP 3389 rapidly exceed 15 attempts per minute, whereas legitimate administrators correcting a password or mistyping credentials generate far fewer failures, suppressing routine noise. |
| Per Target Account | 6–10 failed logons in 1 minute across sources (or hitting 50% of lockout threshold) | Isolates distributed password spraying across rotating source IPs targeting privileged or service accounts, surfacing the anomaly before bad-password counters trigger a denial-of-service account lockout. |
| Host-Wide Aggregate | 60–100 failed logons in 1 minute on a single jump host | Captures low-and-slow distributed sweeps across many source IPs and guessed usernames where individual keys stay below threshold while LSASS.exe authentication work and event log writes remain high. |
For a two-person team, keep the detection and enforcement loop tight and automated. Forward Event ID 4625 records over TLS using a lightweight collector (such as Winlogbeat, NXLog, or Windows Event Forwarding) into a log engine or rule runner that aggregates 1-minute buckets and enforces a 30-to-60-minute suppression window per source IP or account, collapsing a twelve-hour brute-force sweep into a single actionable ticket or webhook. Couple those rate thresholds with automated containment—dynamically inserting temporary deny entries at the perimeter firewall or cloud security group for repeat offending IPs and enforcing Network Level Authentication (NLA) so unauthenticated pre-session probes terminate before full RDP handshake and LSASS.exe logon processing. Relying on manual log review after hours while perimeter ACLs quietly drift guarantees alert fatigue; deterministic burst thresholds paired with automated IP blocks contain the flood mechanically and leave a concise morning summary rather than locked-out accounts.
Edge Cases That Break Simple Burst Rules
- NATted corporate or carrier egress: Many legitimate users sharing a single NAT egress IP can produce transient bursts of failed logons during password rotations or identity/VPN disruptions. Blocking a shared egress IP without accounting for distinct target usernames risks locking out valid staff; pair shared-IP checks with per-account limits and short temporary blocks (15–30 minutes).
- Slow password spraying below the window: An attacker pacing probes at 3 failed logons per minute per account over 10 hours evades a 1-minute burst rule while accumulating 1,800 failures and testing lockout boundaries. Combine short burst buckets with a longer rolling window (such as 25+ failures per account over 1 hour).
- Stale or misconfigured service accounts and scheduled tasks: Scheduled jobs or services authenticating with expired secrets from another host emit recurring 4625 events that mimic low-rate brute-force activity. Explicitly excluding or tagging verified automation identities eliminates recurring false pages while preserving visibility into unexpected account spikes.
- Log forwarding lag or buffer drops during floods: Heavy authentication bursts can stall forwarder queues or saturate local event buffers, delaying alerts or dropping records mid-flood. Monitor forwarder backpressure and event-log drop counters alongside burst rules so log pipeline stalls are never misread as an idle attack surface.
Silent Telemetry: Why Event ID 4625 Failed to Alert
When forty thousand failed RDP logons hit a Windows Server 2019 jump host overnight and nobody wakes up until a human locks out at 8:00 AM, the root failure sits in alert design rather than the Windows Security log. On a two-person IT and security team, paging on every Event ID 4625 floods inboxes within days, driving engineers to mute rules, downgrade severity, or route alerts to queues nobody monitors after hours. Disabling 4625 alerting altogether because isolated failures are routine removes the tripwire entirely: a stale allowlist CIDR can then admit continuous TCP/3389 probes straight into TermService and LSASS.exe without generating a single operational warning.
Once a recycled or reassigned IP prefix remains permitted at the perimeter, external scanners complete a TCP three-way handshake directly against the jump host's network stack. Each connection attempt forces tcpip.sys to hand off RDP negotiation to TermService and CredSSP, NTLM, or Kerberos authentication packages, which invoke LSASS.exe to resolve the target account, validate cryptographic material, update bad-password counters in SAM or domain state, and write a full Event ID 4625 record into the Security.evtx ring buffer before flushing blocks to disk. Across twelve hours, roughly 40,000 attempts average about 55 failures per minute—sustaining enough user-mode context switching, LSASS.exe authentication work, and .evtx write amplification to degrade interactive session responsiveness and lock out valid accounts while leaving the alerting pipeline silent.
Why Event ID 4625 Goes Blind in Lean Environments
Four structural blind spots routinely sever raw authentication telemetry from actionable escalation on small teams:
- Unaggregated single-event rules: Rules triggered by individual 4625 events either page repeatedly on mistyped passwords, expired cached RDP credentials, and routine typos, or get silenced altogether. Without rolling time windows, four hundred attempts in three minutes carry the same rule logic as three stray typos across an afternoon.
- Missing rate correlation across keys: One source IP testing eighty passwords against a single administrative account presents a very different threat profile from sixty source IPs testing two passwords each across thirty accounts. Tracking only host-wide totals or omitting grouping keys buries targeted credential pressure inside routine background noise.
- Absent suppression and cooldown windows: Even burst-based rules generate alert storms during multi-hour brute-force campaigns if they fire on every evaluation interval. Without a 30- to 60-minute per-key cooldown, engineers experience sustained waves as alert noise and lower rule priority.
- Unmonitored collector lag and dropped events: High-rate authentication bursts simultaneously elevate local event generation and upstream log shipping volume. When a forwarder—such as Winlogbeat, NXLog, or Windows Event Forwarding—stalls under TLS backpressure or exhausts local queue buffers, events back up or drop while correlation engines observe zero matching records, turning an active flood into apparent calm.
Burst Thresholds and Alerting Architecture That Won't Bury Two People
For a two-person team operating a Windows Server 2019 jump host, alerting should trigger only when logon failure rates exceed human error patterns on a single host, while collapsing sustained waves into throttled notifications. Rather than relying on host-local scripts or unthrottled rules, ship Event ID 4625 over TLS to a centralized log pipeline or SIEM that evaluates 1-minute and 5-minute rolling windows, enforces a 30- to 60-minute per-key cooldown so an ongoing campaign pages once per window rather than every minute, and couples high-confidence bursts with automated temporary blocks at the edge firewall or security group.
- Per source IP (15 to 25 failed logons in 1 minute, or 40 or more in 5 minutes): A user mistyping a password or reconnecting with a stale saved credential produces a handful of failures, not dozens within sixty seconds. Crossing 15 to 25 failures per minute from one source IP reliably signals automated scanning or credential stuffing, justifying both a cooldown-throttled alert and an automated 15- to 30-minute edge block on that address.
- Per target account (6 to 10 failed logons in 1 minute across sources, or roughly half the account lockout threshold): Attackers rotate source IPs to stay below per-IP thresholds while targeting privileged accounts. Alerting at roughly half the lockout threshold intervenes before an operator locks out mid-incident and before bad-password counter contention compounds.
- Host-wide aggregate (60 to 100 failed logons in 1 minute on the jump host): Even when per-IP and per-account rates stay below individual limits, a distributed sweep across many pairs still drives sustained LSASS.exe authentication work, bad-password counter updates, and .evtx disk writes. A host-level ceiling catches distributed floods before CPU and storage I/O contention degrades interactive jump-host sessions.
Static thresholds still require guardrails against both false positives and slow evasion. Shared NAT or carrier egress gateways can concentrate legitimate failures behind one IP during password resets or identity disruptions, which makes short IP block durations (15 to 30 minutes) and per-account controls far safer than broad subnet bans. Conversely, a low-and-slow password spray at roughly three failures per minute per account over ten hours (1,800 attempts total) stays beneath a 1-minute burst cap while steadily probing accounts; pairing 1-minute burst rules with a longer rolling window—such as 25 or more failures per account per hour—captures slow sprays without paging on normal daily sign-ins. Service accounts and scheduled tasks using expired credentials will likewise trigger per-account rules until inventoried, remediated, or explicitly tagged as verified automation, while enforcing Network Level Authentication (NLA) rejects unauthenticated probes earlier in handshake negotiation before full interactive logon processing runs.
Operational risk accumulates where two slow drifts meet: external IP allowlists decay as address blocks transfer without review, while detection rules decay when teams strip out noisy alerts without replacing them with rate-based detection and automated containment. Remediating both requires treating Event ID 4625 as a rate and capacity problem—capping bursts per IP, account, and host, suppressing repeat alerts during active windows, monitoring collector health so pipeline drops never masquerade as quiet nights, and letting automated perimeter blocks absorb automated hammering before a human ever opens a session.
Anatomy of the Overnight Storm: Five Attack and Failure Scenarios
When a Windows Server 2019 jump host absorbs roughly 40,000 failed RDP logons (Event ID 4625) overnight—averaging 55 failures per minute across a 12-hour window—the primary failure is rarely a missing audit record. Each incoming TCP/3389 probe that clears a stale allowlist completes a three-way handshake in tcpip.sys, hands session negotiation to TermService, and invokes LSASS.exe to resolve security principals and verify credentials. Each rejected attempt increments bad-password counters in SAM or Active Directory state and serializes an Event ID 4625 record into the Security.evtx ring buffer and underlying storage blocks. Under sustained load, user-mode context switching, LSASS.exe cryptographic and authentication work, lock/counter contention on account state, and .evtx write amplification compound on the host while a two-person team sleeps through silence.
On a lean team without a dedicated SOC, alerting on individual Event ID 4625 records leads directly to alert muting because routine typos, expired saved credentials, and mistyped usernames generate continuous background noise. Effective detection requires shifting evaluation out of single-event rules and into a centralized pipeline that ships Event ID 4625 over TLS via Winlogbeat, NXLog, or Windows Event Forwarding into rolling rate buckets with 30-to-60-minute per-key cooldown suppression. Pairing those thresholds with automated temporary perimeter or security-group blocks and Network Level Authentication (NLA) forces unauthenticated probes to fail earlier in CredSSP negotiation before full logon processing executes.
| Aggregation Scope | Burst Threshold | Operational Rationale for a Two-Person Team |
| Per Source IP | 15–25 failed logons / 1 min (or 40+ / 5 min) | Separates automated scanners and credential stuffing from occasional human login errors; pairs with a 15–30 min temporary edge block and 30–60 min alert cooldown. |
| Per Target Account | 6–10 failed logons / 1 min across sources (roughly 50% of lockout threshold) | Detects distributed or IP-rotating brute-force and spray bursts against high-value accounts before bad-password counters reach the lockout threshold and strand active operators. |
| Host-Wide Aggregate | 60–100 failed logons / 1 min on one jump host | Captures multi-source, low-per-key sweeps that still drive LSASS.exe contention, account-state updates, and .evtx disk write amplification. |
1. Stale Allowlist Drift and Uninspected TCP/3389 Ingress
An edge allowlist or security group that retains a reclaimed ISP prefix, a retired VPN egress CIDR, or an oversized subnet mask—such as a /24 or /16 written for a single NAT IP—alters the effective perimeter without modifying the host configuration. When an external scanner sends a TCP SYN to port 3389 on a reassigned or compromised neighbor IP inside that stale range, the packet passes the permissive filter, tcpip.sys returns a SYN-ACK, and the attacker reaches an active RDP listener.
Exposure accumulates through quiet boundary decay: carrier reclamation, cloud IP reallocation, and offboarding rarely trigger reviews of static ingress rules. Once internet-wide scanners probe the recycled address space, every completed handshake pulls external authentication traffic into TermService and LSASS.exe instead of dropping it at the network edge.
2. High-Rate Brute-Force Bursts and LSASS Resource Contention
Once TCP/3389 is reachable, automated credential-stuffing and brute-force tools sustain continuous authentication handshakes against the jump host. Each attempt forces LSASS.exe to resolve the target account, execute credential validation routines, update bad-password state, and emit an Event ID 4625 record to Security.evtx.
At roughly 55 failures per minute over 12 hours, the host sustains repeated user-mode context switches, LSASS.exe authentication cycles, and .evtx write amplification where small log flushes amplify storage I/O overhead. Interactive administrator sessions degrade into lag and delayed input because CPU run queues and disk I/O queues contend with thousands of rejected logon attempts.
3. Account-Lockout Cascades and Shared-Egress Collateral Damage
When attackers target valid usernames or rotate source IPs against shared or privileged accounts, bad-password counters advance toward the configured account lockout threshold. When an administrator or engineer connects the next morning, their authentication attempt hits a locked account state—turning an affected operator into the first indicator of an overnight attack.
Collateral impact intensifies when legitimate staff share a corporate NAT gateway or carrier egress IP during a password rotation or identity/VPN disruption: a broad or long-lived IP ban blocks innocent users sharing that egress address. Calibrating per-account burst alerts to 6–10 failures per minute (roughly half the lockout threshold) and enforcing short 15–30 minute temporary blocks—rather than permanent or wide-CIDR bans—contains automated bursts while limiting collateral lockouts.
4. Alert-Fatigue Suppression and Low-and-Slow Evasion
When a two-person team alerts on individual Event ID 4625 events, routine mistyped passwords and stale RDP session state trigger constant notifications, pushing responders to mute rules or filter them out. Without rolling rate buckets and 30–60 minute per-key cooldowns, sustained overnight floods either overwhelm on-call attention or get suppressed wholesale.
Short-window burst rules alone leave a gap against low-and-slow password spraying: an attacker pacing at roughly 3 failed logons per minute per account over 10 hours (1,800 attempts) stays below a 15–25/min per-IP or 6–10/min per-account 1-minute threshold. Combining 1-minute and 5-minute burst detection with a longer rolling window—such as 25 or more failures per account per hour—and tagging known stale service accounts or scheduled tasks suppresses routine noise without allowing slow sweeps to pass undetected.
5. Collector Backpressure and Overnight Escalation Gaps
High-volume Event ID 4625 generation stresses both the jump host and the telemetry transport path. If TLS sessions stall, forwarder queues in Winlogbeat, NXLog, or Windows Event Forwarding fill up, or local event buffers overwrite under heavy write load, events arrive late or drop out of the SIEM—making an active brute-force run appear as a quiet overnight window.
Pairing telemetry blind spots with no overnight escalation path shifts discovery to morning shift start via a locked-out user report. Tracking forwarder queue depth, lag, and dropped-event counters alongside rate-based burst alerts—while letting automated temporary edge blocks absorb initial volumetric bursts overnight—closes the operational gap between an overnight storm and a morning surprise.
Two-Person Reality: Operational Constraints and Incident Triage
When two generalists split day-to-day infrastructure maintenance, ticket queues, and security triage, pager hygiene governs operational endurance. Any alert that fires on routine post-weekend typos or cached RDP sessions pointing to rotated credentials will be muted within a week. Once single-event Event ID 4625 alerts are silenced, an overnight flood of roughly 40,000 failed logons—approximately 55 attempts per minute over twelve hours—passes unexamined through an unreviewed allowlist gap, surfacing only when morning lockouts stall operations.
The telemetry is present; the detection logic is deficient. Every incoming connection on TCP/3389 completes a three-way handshake in tcpip.sys, hands protocol negotiation to TermService, and invokes LSASS.exe to resolve the target principal, validate credentials, increment bad-password counters in SAM or Active Directory, and commit an audit record through the Security.evtx ring buffer to storage. Sustained overnight at ~55 failures per minute, repeated directory or SAM counter updates, lock contention, context switches, and synchronous or near-synchronous .evtx block flushes increase CPU run-queue pressure and storage I/O latency, degrading interactive admin responsiveness while driving accounts toward lockout.
Catching automated authentication pressure without exhausting a two-person team on routine human errors requires shipping Event ID 4625 over TLS—via Winlogbeat, NXLog, or Windows Event Forwarding—to a centralized collector or SIEM that evaluates rolling rate windows and enforces a 30-to-60-minute per-key cooldown. Instead of emitting one alert per failure record, the rule raises a single incident when a key crosses a burst threshold and suppresses duplicate pages for that key, preventing a multi-hour brute-force campaign from generating continuous mobile alerts.
- Per Source IP (15–25 failures/min, or 40+ failures/5 min): Interactive human retries rarely reach twenty attempts in sixty seconds, whereas automated scanners and credential stuffers routinely exceed this rate. Crossing the threshold triggers a cooldown-gated alert and an automated temporary edge firewall or security-group block for 15–30 minutes, dropping offending packets at the perimeter before they consume LSASS.exe CPU cycles and disk I/O.
- Per Target Account (6–10 failures/min across sources): Set to roughly half of a typical lockout threshold. Distributed botnets rotating source IPs keep per-IP counts below burst limits while driving per-account failures upward; throttling or alerting on this aggregate flags the campaign before bad-password counters lock out privileged or shared administrative accounts outside business hours.
- Host-Wide Aggregate (60–100 failures/min on the jump host): Captures horizontal sweeps sending one or two probes per source IP across many usernames. Even when individual IP and account counters remain modest, a jump-host spike at this rate indicates sustained authentication work and .evtx write amplification capable of slowing host management sessions.
Uniform thresholds fail at operational extremes. Shared NAT or carrier egress pools can push legitimate users past a per-IP ceiling during password-reset waves or identity disruptions; keeping perimeter blocks brief (15–30 minutes) and enforcing per-account limits avoids locking out an entire office or VPN subnet. Conversely, a low-and-slow password spray at ~3 failures per minute per account over ten hours—1,800 attempts in total—stays below a one-minute burst ceiling. Pairing short burst windows with a longer rolling threshold such as 25 or more failures per account per hour catches protracted sprays, while tagging known stale service accounts or scheduled tasks prevents expired credentials from masking active intrusions.
For a two-person team, the most dangerous failure mode is pipeline blindness. During high-rate bursts, TLS handshakes or sends can stall, collector queues can saturate, or local .evtx buffers can overwrite under heavy write load, delaying or dropping Event ID 4625 records so the SIEM appears quiet while brute-force traffic continues hitting TCP/3389. Monitoring forwarder queue depth, ingestion lag, and dropped-event counters alongside automated temporary perimeter blocks prevents a stale allowlist and silent telemetry loss from turning an unmonitored night into a compromised jump host.
Closing the Gap: Hardening Access, Detection, and On-Call Readiness
Forty thousand failed Event ID 4625 logons over twelve hours equals roughly fifty-five failures per minute—not a novel exploit, but sustained mechanical attrition against a drifted perimeter. When a stale CIDR block remains permitted on an edge firewall or cloud security group after a partner office relocates, a VPN egress pool shifts, or an address block is reassigned to another tenant, automated scanners treat TCP/3389 as reachable. Each incoming connection completes a three-way handshake in tcpip.sys, passes protocol handling to TermService, and routes credential verification into LSASS.exe, which resolves account principals, validates secrets against SAM or Active Directory, updates bad-password counters and timestamps, and commits audit records through the Security.evtx ring buffer to disk. Across hours of continuous volume, that execution path generates user-mode context switches, lock contention on account state records, and small-block write amplification on the event log volume, degrading interactive administrative responsiveness while neither rotation member receives a page.
Burst Thresholds and Alerting Architecture for a Lean Team
Single-event 4625 alerts fail within days on a two-person on-call rotation. Routine mistyped passwords, cached RDP sessions carrying expired credentials, and legacy scheduled tasks produce continuous background volume until an engineer suppresses or mutes the alert—leaving the jump host unmonitored when a genuine brute-force flood begins. Rather than paging on isolated failures, forward Event ID 4625 over mutual TLS (via Winlogbeat, NXLog, or Windows Event Forwarding) into a centralized log collector or SIEM that evaluates sliding windows and applies a 30- to 60-minute per-key cooldown, collapsing a prolonged credential attack into one actionable incident instead of repeated pages.
- Per source IP (15–25 failures/minute, or 40+ failures/5 minutes): Isolates automated credential stuffing and scanner loops from normal human retry patterns, and triggers a 15- to 30-minute perimeter firewall or security-group drop so hostile packets are discarded before reaching LSASS.exe or storage I/O.
- Per target account (6–10 failures/minute across sources): Triggers well below standard account-lockout thresholds so distributed, IP-rotating attempts against privileged or shared accounts surface before lockouts lock out operators off-hours.
- Host-wide aggregate (60–100 failures/minute on the jump host): Captures broad username sweeps whose per-IP and per-account rates stay low enough to evade narrow keys while still driving LSASS.exe CPU load, counter updates, and .evtx flush pressure across the host.
Rate-based detection fails silently if the telemetry pipeline saturates during the burst. Under heavy authentication and audit load, TLS handshakes or streams can stall, local forwarder queues can fill, and circular .evtx buffers can overwrite unread records faster than collectors drain them, producing a quiet SIEM dashboard while TCP/3389 is under heavy load. Monitor forwarder queue depth, ingestion lag, and dropped-event counters alongside the burst rules so telemetry degradation pages the on-call engineer rather than masking an active intrusion.
Edge Cases, Evasion, and Actionable Defenses
Fixed rate thresholds face clear trade-offs when adversaries throttle their pace or legitimate users share egress infrastructure. A low-and-slow password spray averaging three failures per minute per account over ten hours (1,800 attempts) stays below a one-minute threshold, requiring a longer rolling window—such as 25 or more failures per account per hour—paired with explicit tagging of retired service accounts and scheduled tasks so stale automation does not obscure real attacks. Conversely, shared NAT gateways, carrier-grade NAT pools, and centralized VPN exits can push legitimate administrators past a per-IP threshold during password rotations or identity outages; keeping automated edge blocks brief (15–30 minutes) and prohibiting broad /24 or /16 automated bans prevents collateral disruption across unrelated users.
- Eliminate direct internet exposure: Place jump-host access behind a VPN or identity-aware zero-trust broker enforcing MFA, require Network Level Authentication (NLA) so unauthenticated probes fail earlier during CredSSP negotiation rather than invoking full logon routines, and enforce periodic allowlist reviews tied to an accountable owner.
- Automate containment at the edge: Connect high-confidence source-IP burst detections directly to perimeter firewall or security-group temporary blocks so offending traffic drops upstream of LSASS.exe contention and disk write amplification.
- Keep on-call playbooks tight and testable: Standardize escalation steps for a two-person rotation around verifying allowlist drift, checking forwarder health and dropped-event counters, resolving or investigating lockouts, and rotating compromised credentials—eliminating pages for single failed logons.
Top comments (0)