Detectors require sufficient captured session history before they can produce eligible evaluations. A finding list that remains empty does not mean collection is broken — it may mean the org does not yet have enough history for a given detector to fire.
How anomaly detection works
Each detector runs once per completed daily window, per organization. For each subject it evaluates (a member, a session, or an operation), the detector produces one of:
When a finding fires repeatedly, it updates the same open finding record rather than creating duplicates. Administrators can acknowledge, resolve, or dismiss each finding through the Signals tab.
Detector summary
Coverage concepts and the admin UI are documented in Coverage Alerts and the Coverage tab.
AD-001 — Usage baseline deviation
What it checks: Whether a member’s completed session or token volume on a given day is materially above their own recent history. How it is calculated: RCLM looks at the previous 14 completed UTC days for that member, requires at least 7 active baseline days, and builds a guarded exponentially weighted baseline. A finding requires all three of the following to align:- A strong statistical deviation (z-score above threshold)
- A meaningful ratio above the member’s own baseline
- An absolute increase floor (5 sessions or 100k tokens, depending on the variant)
AD-002 — New technology fingerprint
What it checks: Whether a session contains a first-seen combination of dependencies, imports, manifests, tools, providers, or project activity for that member, compared against their prior 90-day history. How it is calculated: RCLM builds a fingerprint from each session’s file extensions, package manifests, conservatively parsed dependencies and imports, command families, tools, models, providers, agent clients, and project roots. A single new item is not enough — the detector requires a meaningful combination, such as:- A first-seen package ecosystem together with a related first-seen dependency.
- A new project root combined with a new command family.
AD-003 — Runaway or looping session
What it checks: Whether a session shows signs of being stuck — through repeated identical failures, failed build or test loops, abnormal tool-call volume, or abnormal duration. How it is calculated: The detector applies deterministic sub-signals:
Severity escalates by rule:
- Warning — one sub-signal crossed.
- High — two independent sub-signals, or a sustained failed build/test loop.
- Token volume alone never produces a critical result.
AD-004 — Sensitive file access
What it checks: Whether sessions in the window read, wrote, or referenced.env-class files, credential files, or private-key files — including cases where the AI agent wrote a script or temporary file that reads or modifies those files.
How it is calculated: RCLM scans three distinct surfaces per session:
Surface 1 — Direct tool access
Each tool call’s file paths are checked against sensitive-file patterns:.env,.env.*,*.env,.envrc*.pem,*_rsa,*id_rsa,*.p12*credentials*.json,*service-account*.json
primary_file_path and file_paths[] on native read and write tool calls. RCLM also records whether hook DLP was active (was_redacted), so evidence distinguishes covered accesses from DLP gaps.
Surface 2 — Shell command access
File paths extracted from shell commands (the same extraction used by the P2 Groundhog Files signal):cat .env.production,head .envrc,tail .env.localsed -n 'p' dev.env,rg "SECRET" .env
Surface 3 — Code and script mentions
The agent writes a Python, Bash, or other script that opens, loads, or references a sensitive file — including temporary scripts written to disk and immediately executed. Examples:load_dotenv() calls, source .env-style shell expressions, and similar constructs. A mention in agent-written code is recorded as a distinct access type (code_mention) separate from a direct tool read or shell command.
This surface matters because hook DLP operates on tool inputs and outputs at the boundary. A model that writes a script referencing an env file as a string does not pass the secret through a hook — the script itself becomes the exfiltration vector when it is later executed. AD-004 surfaces this so the admin can decide whether that script was actually run and whether the execution context had DLP coverage.
Severity tiers:
- Warning — one or more sessions with a sensitive file read or
code_mention. - High — at least one session wrote to a sensitive file path, or the code mention involves a write operation (
open(".env", "w"), overwriting viadotenv_values). - Critical — multiple sessions have env-file activity and at least one had no DLP coverage, indicating a gap where secrets or the script referencing them may have reached an external model provider.
read, write, shell_read, code_mention), matched file pattern categories, DLP coverage count per session, and session IDs.
Evidence never includes: Actual file paths, file contents, secret values, script source code, or redacted placeholder text.
Interpretation guardrail: Accessing or referencing env files is a normal part of debugging and configuration work. A code_mention finding means the agent wrote code that references a sensitive file — not that the secret was extracted. Whether secrets were exposed depends on whether the script was executed and whether hook DLP covered the execution context.
Typical next steps:
- Check the access type in evidence. A
code_mentionwarrants a different response than a directreadwithwas_redacted: false. - For
code_mention, open the linked session and review the script the agent wrote. Determine whether it was executed and whether that execution happened within a DLP-covered hook boundary. - For direct reads or shell reads with a DLP gap (
was_redacted: false), determine whether secrets were transmitted to an external provider. - If the session used Gemini, Antigravity, or a capture-only integration — clients that do not support pre-execution input redirection — treat any sensitive file activity as a potential gap regardless of
was_redacted. - Open Files for the project and window. Sensitive Only shows the filename matches, and Sensitive Reviews is where an admin records clear or flag.
Hook DLP default behavior:
rclm-hooks-install enables DLP on fresh installs. Users who opted out with --no-dlp or installed before the default changed may have sessions with no DLP coverage. AD-004 Critical findings are the primary signal for this gap.AD-005 — Proxy operation volume spike
What it checks: Whether a canonical proxy operation’s daily session volume is materially above the organization’s own recent daily history for that operation. How it is calculated: RCLM groups proxy sessions by their canonical operation title. For the latest completed day, it compares that operation’s session count against its own active-day history (up to 90 days, requiring at least 7 active baseline days). A finding requires all three of:- A strong statistical deviation (z-score above threshold)
- A meaningful ratio above the operation’s own baseline
- An absolute increase floor (5 sessions)
AD-006 — New proxy operation
What it checks: Whether a canonical proxy operation title appears for the first time in the organization’s recent proxy session history. How it is calculated: RCLM compares the canonical operation titles seen in the latest completed day against the preceding 90 days of labeled proxy activity. At least 10 prior labeled proxy sessions are required before a title is treated as genuinely new. A new hash mapped to an existing canonical title is not treated as a new operation. Severity: Warning. Evidence includes: The canonical operation title, first-seen date, and linked session IDs. Evidence never includes: Request bodies, prompt content, or credentials. Interpretation guardrail: A first-seen operation title is not inherently unsafe. New integrations, API migrations, and approved workflow changes all produce this finding. Typical next steps: Verify that the new operation belongs to an approved integration or workflow. If the title is unrecognized, inspect the linked proxy sessions to identify the caller before deciding whether to escalate.AD-010 — Hooks tampered or missing
What it checks: Whether a coding agent’s ReclaimLLM hooks leave a healthyinstalled state after the device’s first baseline observation.
How it is calculated: Device check-ins upsert per-client status. A finding fires when status moves from installed to missing, partial, disabled, or modified; when policy_signature_valid is false; or when the client sends event: uninstall. Transitions within 30 minutes of an rclm version change are held (upgrade/reinstall grace). The first observation of a client is baseline only — it does not fire AD-010.
Severity: High if the user had sessions on that client in the last 14 days; otherwise warning-lane medium.
Evidence includes: Client name, from/to status, device id, and timing metadata.
Evidence never includes: Hook file contents, hostnames in cleartext, or secrets.
Interpretation guardrail: Local-only installs count as healthy when cwd-local hooks are present. Alternate config directories and unmanaged clients can look missing without malice.
Typical next steps: Ask the developer to run rclm-hooks-install --repair, or confirm an intentional uninstall / alternate HOME. Findings resolve when the client returns to installed.
AD-011 — DLP disabled
What it checks: Whether env-file DLP is off on a device without an active exception, or whether a developer attempted--no-dlp while the org requires DLP.
How it is calculated: Check-ins report resolved dlp.enabled, local setting, and whether org dlp_required applies. Session policy snapshots can also surface DLP-off for users without a recent check-in. Active DLP exceptions suppress the finding until expiry.
Severity: High when DLP is required and still disabled (implies override/tamper); otherwise medium. Attempted opt-out while required may be informational.
Evidence includes: DLP flags, exception id when present, device id.
Interpretation guardrail: Local --no-dlp is allowed until an admin turns on Require DLP. Grant time-boxed exceptions for legitimate cases.
Typical next steps: Enable Require DLP in Coverage settings, grant an exception, or ask the developer to reinstall without --no-dlp.
AD-012 — Device silent
What it checks: Whether a formerly active durable device stops reporting while the user still appears active elsewhere. How it is calculated: Baseline requires enough recent active days (session or check-in). Silence means no sessions, check-ins, or bootstraplast_seen_at for the configured business-day threshold. Medium severity requires corroboration (gateway proxy records, activity on another device, or web login). Without corroboration the finding stays informational and is not emailed. Ephemeral devices and retired devices are suppressed.
Severity: Warning-lane medium with corroboration; info without.
Interpretation guardrail: Laptops off for PTO are expected. Retire unused devices. Silence alone does not prove tamper.
Typical next steps: Confirm the machine is offline or retired, or ask for rclm-hooks-install --repair / rclm-heartbeat.
AD-013a — Gateway without hooks
What it checks: Whether a user generates gateway (proxy) traffic without corresponding hook-captured sessions. How it is calculated: Daily comparison ofrecord_type=proxy volume vs hook session records in the evaluation window. Inactive when the org has no gateway configuration.
Severity: Warning-lane medium.
Typical next steps: Confirm the developer still has hooks installed, or that traffic is expected from a non-hook path. See LLM Gateway.
AD-013b — Direct provider routing
What it checks: Whether check-in routing classification reportsdirect for a provider the org marked as required to use the gateway.
How it is calculated: SessionStart routing probe classifies env/base-URL posture into gateway / direct / unknown without sending URLs or keys. Only direct for configured gateway_required_providers fires. unknown never counts as bypass.
Severity: Warning-lane medium.
Typical next steps: Point the agent at the org gateway hosts, or adjust required providers if the agent cannot use the gateway (for example some IDE-hosted models).
AD-013c — Provider usage unreconciled (Enterprise)
What it checks: Whether provider admin usage reports diverge from gateway-attributed tokens for the same day. How it is calculated: A background job decrypts Enterprise reconciliation credentials, fetches provider usage (best-effort), and compares against gateway ingest. Designed with a lag for provider API delay. Flag usage on keys outside org credentials or provider tokens substantially above gateway totals. Severity: High when reconciliation fires. Interpretation guardrail: Personal provider accounts and non-gateway paths are invisible to this check. Treat results as investigation leads, not perfect metering. Typical next steps: Rotate unexpected keys, confirm gateway enrollment, or disable reconciliation if credentials are stale. How to produce AD-010 through AD-013c in the product is in Signals → Coverage and tamper.AD-015 — External network egress
What it checks: Sessions where an AI coding agent initiates an outbound network connection to a non-local destination across any model or coding client. Surfaces inspected:
Local loopback targets (
localhost, *.localhost, 127.0.0.0/8, ::1, 0.0.0.0, host.docker.internal) are automatically excluded so routine local dev-server checks do not fire findings.
Severity tiers:
- Warning: Read-only native
WebFetchtool calls to standard external HTTPS domains with no shell, transfer, or upload behavior. - High: Any shell HTTP (
curl,wget), shell transfer/socket tool (scp,rsync,nc), programmatic script egress,/tmpstaging download (-o /tmp/...or> /tmp/...), or direct connection to a raw external IP literal. - Critical: Any of the following high-risk indicators:
- Outbound payload upload or exfiltration (
POST,PUT,PATCH,--data*,-d,-F,--upload-file,-T, outboundscp/rsync/sftp, or scriptrequests.post/requests.put/urllibdata upload) - Cloud instance metadata SSRF probes (
169.254.169.254,100.100.100.200,fd00:ec2::254,metadata.google.internal) - Remote code execution chains (
curl ... | sh,wget -O- ... | bash) - Egress occurring in the same session as sensitive file access (
.env,.ssh, private keys, cloud credentials) or system/DNS reconnaissance (/etc/resolv.conf,/etc/hosts,uname -a,whoami,id)
- Outbound payload upload or exfiltration (
AD-015:
- In Signals → Configure Email Alerts & Cadence → Organization (or Settings → Email Alerts & Cadence), add entries under Allowed Egress URLs, Domains & IPs (AD-015 Safe List).
- Accepted formats: full URLs (
https://api.example.com/v1— normalized to hostname), domains (github.com— matchesgithub.comand subdomains likeapi.github.com), wildcards (*.internal.corp), IPv4/IPv6 addresses (203.0.113.10), and CIDR ranges (10.20.0.0/16). - On any
AD-015finding card in Anomaly reports, admins can also click Mark destination safe to add the finding’sdestination_hosts/destination_ipsto the organization safe list and dismiss the finding in one step. - Security guardrail: Cloud metadata SSRF addresses (
169.254.169.254,metadata.google.internal,169.254.0.0/16) can never be added to the safe list.
destination_hosts (hostname only — never URL paths, query strings, or auth headers), destination_ips, access_types, egress_methods, command_families, model, egress_ToolCallCount, and boolean flags (sensitive_file_chain, system_recon_chain, metadata_ip_probe, pipe_to_shell, tmp_Drop).
Interpretation guardrail: Outbound network calls can occur during legitimate API testing, documentation lookups, or package downloads. Check the destination host/IP, egress method, and whether the call was read-only (GET) or an upload (POST/PUT/scp) before escalating.
Typical next steps:
- Inspect
destination_hosts,destination_ips,egress_methods, andmodelon the finding card. - Open the linked session to verify whether the network call was user-directed or autonomous.
- If the destination is an approved corporate service or registry, click Mark destination safe (or add the domain/IP/CIDR in Alert Settings).
- If the finding is Critical (upload, metadata probe,
curl | sh, or chained with.env/credential reads), rotate any potentially exposed secrets and investigate the workstation.
AD-016 — PII content exposure
What it checks: Whether a session’s content — a proxy request/response body or a hook/agent tool-call input/result — contains a validated email address, phone number, Social Security Number, or credit card number. How it is calculated: RCLM scans two surfaces, depending on how the session was captured:- Hook/agent sessions: tool-call input and result previews.
- Gateway proxy traffic: the request/response body preview captured at ingest.
- Credit card candidates are confirmed with a Luhn checksum.
- SSN candidates are checked against Social Security Administration invalid-range rules (certain area, group, and serial number patterns are never issued).
- Phone candidates exclude reserved fictional test ranges.
- Email matches a standard address pattern.
- Proxy traffic: PII here has already reached the LLM provider’s API — that is what a proxy record is. Severity starts at High and escalates to Critical once the same member has matches across multiple sessions.
- Hook/agent sessions: a local tool-call read or result has not left the machine yet. Severity starts at Warning, rises to High with multiple categories or multiple sessions, and only reaches Critical when the same session also has an AD-015 external network egress finding — meaning the data was both touched locally and sent somewhere external in the same session.
email, phone, ssn, credit_card), per-surface session counts (proxy vs. hook/agent), a boolean flag for whether the finding is chained with an AD-015 egress finding, and a bounded list of session IDs.
Evidence never includes: The actual email address, phone number, SSN, card number, or any other matched value — only the category label and count are recorded.
Interpretation guardrail: A proxy finding means the flagged data already reached the model provider; treat it as having left the organization’s perimeter. A hook/agent finding without an egress chain means the data was read or produced locally and has not (yet, as far as RCLM can see) left the machine — the severity reflects that difference.
Typical next steps:
- Check whether the finding is a proxy or hook/agent match in the evidence — this changes what already happened versus what might happen next.
- For a proxy finding, confirm whether the destination provider is an approved one, and whether the workflow that sent this data should be changed to avoid sending it at all.
- For a hook/agent finding, check
chained_with_egress. If true, treat it the same as an AD-015 Critical finding — the data was both touched and sent externally in one session. - Open the linked session. On the session detail page, matching tool calls and messages are automatically expanded and highlighted so you can see exactly where the match occurred without it being hidden behind a collapsed view.
AD-017 — Proxy failure rate
What it checks: Whether the share of Enterprise LLM Gateway calls that fail is materially above the organization’s own recent failure rate. It covers the gateway only; calls through the local proxy are not scored. How it is calculated: For each completed hour and each completed day, RCLM computes:failure rate = (error responses + interrupted streams) ÷ calls with a known status
- An error response is any status other than 2xx, including provider timeouts and provider rate-limit or out-of-credit responses.
- An interrupted stream is a streamed response that failed after it had started. The client already received status 200, so RCLM records it as a 200 with a stream-interrupted flag and counts it as a failure.
- Calls with no recorded status are never counted as failures and are removed from the denominator.
- At least the minimum number of calls (default 20) and failures (default 5) in the window.
- A failure rate at least 2× the organization’s usual rate, and at least 5 percentage points higher.
- A strong statistical deviation (z-score above threshold).
- Operation subjects return insufficient data while some calls in the window are still waiting for an operation label. The organization-wide rate is always evaluated, so a labeling delay cannot hide an outage.
- A separate capture-health finding fires when 10% or more of calls in a window have no recorded status.
- Check whether the finding is organization-wide or limited to one operation. An organization-wide rise usually points to a provider outage, an expired provider credential, or exhausted provider credits.
- Open the linked sessions and compare status codes and error classes. Many 429s indicate a provider rate limit; many timeouts indicate a slow or unreachable provider.
- For an operation-only rise, look at what changed for that workload: model, prompt size, or caller.
AD-018 — Proxy latency regression
What it checks: Whether a model has become materially slower on the Enterprise LLM Gateway compared with its own recent history. How it is calculated: Once per completed day, for each model, RCLM looks only at successful, non-streaming calls. It compares the day’s 95th-percentile latency (p95) and median latency with that model’s own earlier days (up to 14) that had enough calls. A finding requires all of:- At least the minimum number of successful calls that day (default 20).
- At least the required number of earlier days with enough calls (default 3). Until then the model is not scored.
- p95 at least 1.5× the usual p95, at least 500 ms slower, and a strong statistical deviation.
- Median at least 1.25× the usual median.
Tuning AD-017 and AD-018
Organization administrators can adjust the main thresholds for both detectors in Signals → Proxy Thresholds (AD-017/018), next to Egress Policy (AD-015). Each field shows its default; leave it blank to use the default. Values outside the allowed range are rejected when you save.
Lower values make a detector more sensitive and produce more findings; higher values reduce noise. The z-score requirement, the baseline length, and the 5× baseline High rule are fixed.
Evidence design
All detectors follow the same evidence boundary: findings store compact, allowlisted metadata — identifiers, normalized hashes, counts, categories, statistics, and thresholds. They never store raw prompts, tool inputs, tool results, file contents, secrets, complete shell commands, or personal absolute paths. This means anomaly detection does not create a second sensitive-data problem while providing enough structured information to investigate each finding.Finding lifecycle
Each open finding has a single lifecycle record per detector, subject, and anomaly target. Repeated detections update the last-seen time and occurrence count rather than creating duplicate entries.
Resolved and dismissed findings are closed. Missing data or a detector error cannot silently resolve an open finding.
Email alerts for anomalies
After detector windows complete, RCLM can email acute findings as immediate P0 alerts and roll others into daily or weekly digests. P0 covers AD-004 Critical (or High with write access), AD-003 High/Critical, AD-015 High/Critical, AD-016 Critical, and the organization-wide AD-017 High failure-rate finding. When a dedicated Security inbox is configured, AD-004 Critical and AD-015 Critical always route to that inbox (unless every destination on anAD-015 finding is covered by the org’s allowed egress safe list). Digests run at 08:00 recipient-local time.
Email payloads stay inside the same privacy boundary as finding evidence: categorical labels and counts only — never titles, paths, commands, or prompts. Configure routing and personal preferences from Signals Dashboard → Email alerts & notification cadence.