Deception Check

FIELD RESEARCH / AI TOOLING AND WEB EXPOSURE / NOVEMBER 1, 2026

Scanners are asking for your AI-tool files.

Your assistant's configuration belongs in your development workflow. Check whether a copy has made it into the files your website serves.

The scanner's dictionary now includes AI assistants, coding tools, and the files that connect them to other services.

Our decoys recorded 1,429 requests for AI-tooling paths in a retained month of configuration-file hunting. Twenty-one asked for /.claude/.credentials.json. Visitors were trying the names of development files against a public web server. Those requests show what they sought, not that they retrieved a working secret.

You do not need to run a decoy to use this finding. Pick one public application and ask its owner a concrete question: what would our server return for these paths?

Open the files you actually ship

A release artifact is the package your deployment uses: a folder, archive, container image, or built website. It may differ from the source repository. Check the last package before it reaches the public server, then check the routes the server exposes.

DEPLOYMENT BOUNDARY / FOUR PRACTICAL CHECKS

  1. 01 / WORKSPACEKnow what is private

    Inventory local credentials, connection settings, and tool configuration.

  2. 02 / BUILDInspect what you ship

    Check the final directory, archive, or image filesystem, not just the source repository.

  3. 03 / SERVINGCheck the public route

    Confirm the web root, file rules, CDN, and cache behavior.

  4. 04 / RESPONSERead what comes back

    A status code alone cannot tell a normal page from an exposed file.

This is a recommended review workflow, not a reconstruction of an observed exposure. A private local file is not automatically a public web resource.

A .gitignore entry controls what Git tracks; it does not define everything your deployment copies. For Docker builds, Docker documents a separate .dockerignore file that removes matching files from the build context. Use an explicit list of files to publish, then inspect the resulting artifact. If you need a secret during a build, use the builder's supported secret mechanism rather than copying it into the release.

For the person approving the release: Ask for three receipts: the artifact that was checked, the external response that was reviewed, and the name of the person who owns any fix.

The requested filenames tell a useful story

SEPTEMBER 8 TO OCTOBER 7, 2026 / SELECTED CONFIGURATION-PATH EXPORT

2,317 request records
1,429AI-tooling paths
888Other selected paths
The denominator is this selected configuration-path export, not all traffic on the internet or all traffic to our decoys.

The requests named Claude settings, Cursor and Continue configuration, MCP files, provider-named configuration, and LiteLLM configuration. MCP, the Model Context Protocol, lets an assistant connect to tools and data sources. Its configuration may describe servers, commands, or authentication settings.

ONE EXACT REQUESTED PATH / RETAINED EVIDENCE
GET /.claude/.credentials.json
21 exact-path requests in the same window.

A readable path-and-method excerpt, not a complete request replay. Source addresses and credential contents are not displayed.

These names do not all mean the same thing. Some files can hold authentication material. Others contain ordinary project settings or connection details. Some are intended to be shared inside a repository. Inspect the actual contents and intended audience before deciding what needs protection.

What the filenames mean today
Tool or fileUseful contextWhat to check
Claude Code credentialsAnthropic documents .claude/.credentials.json storage on Linux and Windows, with a macOS fallback when Keychain cannot store a login.Keep authentication material outside the public release.
Cursor MCP configurationCursor documents project and user .cursor/mcp.json locations.Review embedded environment values, headers, commands, and server addresses.
Continue configurationContinue uses .continue/config.yaml; config.json is a legacy format.Review both current and legacy files in old artifacts.
Other requested namesA scanner can guess a plausible filename even when no current tool uses that location.Use observed names as hunt leads, not an authoritative configuration inventory.

Separate a request from an exposure

An HTTP 200 means the server treated a request as successful. It does not tell you what the body contained. A single-page application may return its normal HTML page for every unknown route. A cache may return a copy the origin no longer serves.

REQUEST

Someone asked

Keep the raw path, time, method, available source context, and nearby requests. Review approved tests separately.

RESPONSE

What did we serve?

Compare content type, size, cache behavior, and the actual response with the expected page. Keep sensitive content restricted.

USE

Was a secret used?

If real authentication material was exposed, preserve the evidence, revoke it promptly, and review the provider's access records.

If you confirm an exposed key, removing the file is only one part of the fix. Follow the provider's revocation or rotation process, investigate its use, and correct the copy or serving rule that made it reachable. OWASP's secrets-management guidance covers secret lifecycle, access control, and response to exposure.

For monitoring, keep failed requests as reconnaissance context and prioritize confirmed sensitive content separately. Group a dictionary sweep into one investigation with its original records attached. Do not turn every guessed filename into a separate emergency.

Detection appendix: ordinary web logs, useful questions

Start with the Sigma rules below. They use conventional web-server request fields. The pack also includes Microsoft Sentinel IIS, Splunk, Sumo Logic, Elastic, and CrowdStrike CQL access-log hunts, plus local artifact checks. None requires Deception Check telemetry.

Safe-fixture and source checks are recorded in the pack. Native searches still need the declared ingestion, supported functions, and a test in your environment; native query compilation and customer-tenant execution are not claimed. A request match is a triage lead, not an exposure verdict.

Sigma / web access logs

Request for a Claude Credential File

Identify requests for the selected Claude credential path.

Requires: Webserver access log; cs-uri-stem mapped to the path only; SIEM conversion pipeline.

title: Request for a Claude Credential File
id: 8dc24394-17b2-4b3d-83fc-238e802e648a
status: experimental
description: >-
    Detects an HTTP request whose path ends with the selected Claude credential
    filename. A match establishes a request, not file existence or retrieval.
references:
    - https://code.claude.com/docs/en/authentication
    - https://sigmahq.io/sigma-specification/specification/sigma-appendix-taxonomy.html
author: Deception Check Research
date: 2026-11-01
logsource:
    category: webserver
    definition: Access logs with cs-uri-stem mapped to the request path without a query string. No URL decoding is performed by this rule.
detection:
    selection:
        cs-uri-stem|endswith: '/.claude/.credentials.json'
    condition: selection
fields:
    - c-ip
    - cs-host
    - cs-method
    - cs-uri-stem
    - sc-status
falsepositives:
    - Authorized exposure assessments or intentionally monitored decoys
    - Applications intentionally serving an identically named resource
level: medium

Sigma / web access logs

Request for Selected AI Tool Configuration Files

Identify requests for nine selected configuration paths.

Requires: Webserver access log; cs-uri-stem mapped to the path only; SIEM conversion pipeline.

title: Request for Selected AI Tool Configuration Files
id: e6249c30-e9d9-4079-a626-c81c8cb0a057
status: experimental
description: >-
    Detects HTTP requests for selected Claude, Cursor, Continue, and MCP
    configuration paths. Some configurations are intentionally shared and may
    contain no credentials. A match establishes a request, not exposure.
references:
    - https://code.claude.com/docs/en/settings
    - https://code.claude.com/docs/en/mcp
    - https://prod.cursor.com/help/customization/mcp
    - https://docs.continue.dev/customize/deep-dives/configuration
    - https://modelcontextprotocol.io/docs/2026-07-28/develop/connect-local-servers
author: Deception Check Research
date: 2026-11-01
logsource:
    category: webserver
    definition: Access logs with cs-uri-stem mapped to the request path without a query string. No URL decoding is performed by this rule.
detection:
    selection:
        cs-uri-stem|endswith:
            - '/.claude/settings.json'
            - '/.claude/settings.local.json'
            - '/.claude.json'
            - '/claude_desktop_config.json'
            - '/.mcp.json'
            - '/.cursor/mcp.json'
            - '/.continue/config.yaml'
            - '/.continue/config.json'
            - '/.continuerc.json'
    condition: selection
fields:
    - c-ip
    - cs-host
    - cs-method
    - cs-uri-stem
    - sc-status
falsepositives:
    - Authorized exposure assessments or intentionally monitored decoys
    - Intentionally published examples or shared project configuration
level: low

Choose the native hunt that matches your logs

Use the Microsoft query for IIS logs, Splunk and Sumo Logic for their declared access-log sources, and Elastic for its Nginx integration. The CrowdStrike CQL version searches Apache access logs ingested into Falcon LogScale or Next-Gen SIEM. A Falcon endpoint sensor alone does not supply those inbound HTTP paths. Each guide names the actual collection route and query prerequisites.

Microsoft Sentinel / Azure Monitor

AI file requests in native IIS logs

Hunt both selected path groups with one decode pass.

Requires: IIS W3C logs in W3CIISLog through Azure Monitor Agent and an IIS data collection rule; query permission.

// Hunt starter for native W3CIISLog. See README.md before use.
// Manually authored translation of both Sigma path lists, with one decode pass.
// A match is a request, including 404; status does not establish a served secret.
W3CIISLog
| where TimeGenerated >= ago(24h)
| where isnotempty(csUriStem)
| extend OriginalPath = tostring(csUriStem)
// Split BEFORE decoding so a query-value filename cannot become a path match.
| extend MatchPath = tolower(url_decode(tostring(split(OriginalPath, "?")[0])))
| extend CredentialPath = MatchPath endswith "/.claude/.credentials.json"
| where CredentialPath
    or MatchPath endswith "/.claude/settings.json"
    or MatchPath endswith "/.claude/settings.local.json"
    or MatchPath endswith "/.claude.json"
    or MatchPath endswith "/claude_desktop_config.json"
    or MatchPath endswith "/.mcp.json"
    or MatchPath endswith "/.cursor/mcp.json"
    or MatchPath endswith "/.continue/config.yaml"
    or MatchPath endswith "/.continue/config.json"
    or MatchPath endswith "/.continuerc.json"
| extend Finding = iff(CredentialPath, "credential-path request", "configuration-path request")
| project TimeGenerated, Computer, csHost, cIP, csMethod, OriginalPath, MatchPath, scStatus, scBytes, Finding
| order by TimeGenerated desc

Splunk Enterprise / Cloud

AI file requests in combined access logs

Hunt both selected path groups with one decode pass.

Requires: Replace YOUR_WEB_INDEX; access_combined source; validated uri/clientip extraction; search permission.

index=YOUR_WEB_INDEX sourcetype=access_combined earliest=-24h
| where isnotnull(uri) AND len(uri)>0
| eval OriginalTarget=uri
| eval MatchPath=lower(urldecode(mvindex(split(OriginalTarget,"?"),0)))
| eval CredentialPath=if(match(MatchPath,"/[.]claude/[.]credentials[.]json$"),1,0)
| where CredentialPath=1 OR match(MatchPath,"/([.]claude/settings[.]json|[.]claude/settings[.]local[.]json|[.]claude[.]json|claude_desktop_config[.]json|[.]mcp[.]json|[.]cursor/mcp[.]json|[.]continue/config[.]yaml|[.]continue/config[.]json|[.]continuerc[.]json)$")
| eval Finding=if(CredentialPath=1,"credential-path request","configuration-path request")
| table _time host source clientip method OriginalTarget MatchPath status bytes Finding
| sort 0 - _time

Sumo Logic Log Search

AI file requests in Sumo Logic Apache logs

Hunt both selected path groups in inbound access logs.

Requires: Apache Common/Combined Log Format raw logs; replace source category; Log Search permission.

_sourceCategory="YOUR_APACHE_ACCESS_CATEGORY"
| parse regex "^(?<hunt_client>[^ ]+) [^ ]+ [^ ]+ \\[[^\\]]+\\] \"(?<hunt_method>[^ ]+) (?<hunt_target>[^ \"]+) HTTP/[0-9.]+\""
| parse regex "^(?:[^ ]+) [^ ]+ [^ ]+ \\[[^\\]]+\\] \"(?:[^ ]+) (?:[^ \"]+) HTTP/[0-9.]+\" (?<hunt_status>[0-9]{3})(?: |$)" nodrop
| parse regex field=hunt_target "^(?<hunt_raw_path>/[^?]*)"
| urldecode(hunt_raw_path) as hunt_decoded_path
| toLowerCase(hunt_decoded_path) as hunt_path
| where hunt_path matches "*/.claude/.credentials.json"
    or hunt_path matches "*/.claude/settings.json"
    or hunt_path matches "*/.claude/settings.local.json"
    or hunt_path matches "*/.claude.json"
    or hunt_path matches "*/claude_desktop_config.json"
    or hunt_path matches "*/.mcp.json"
    or hunt_path matches "*/.cursor/mcp.json"
    or hunt_path matches "*/.continue/config.yaml"
    or hunt_path matches "*/.continue/config.json"
    or hunt_path matches "*/.continuerc.json"
| if(hunt_path matches "*/.claude/.credentials.json", "credential", "configuration") as hunt_group

Elastic ES|QL 9.2+

AI file requests in Elastic Nginx logs, decoded

Hunt both selected path groups in inbound access logs.

Requires: Nginx integration access streams; single-valued ECS url.original; read and ES|QL permission.

FROM logs-nginx.access-*
| WHERE event.dataset == "nginx.access"
| EVAL hunt_original_target = url.original
| WHERE hunt_original_target IS NOT NULL
| EVAL hunt_raw_path = MV_FIRST(SPLIT(hunt_original_target, "?"))
| WHERE STARTS_WITH(hunt_raw_path, "/")
| EVAL hunt_path = TO_LOWER(URL_DECODE(hunt_raw_path))
| WHERE ENDS_WITH(hunt_path, "/.claude/.credentials.json")
    OR ENDS_WITH(hunt_path, "/.claude/settings.json")
    OR ENDS_WITH(hunt_path, "/.claude/settings.local.json")
    OR ENDS_WITH(hunt_path, "/.claude.json")
    OR ENDS_WITH(hunt_path, "/claude_desktop_config.json")
    OR ENDS_WITH(hunt_path, "/.mcp.json")
    OR ENDS_WITH(hunt_path, "/.cursor/mcp.json")
    OR ENDS_WITH(hunt_path, "/.continue/config.yaml")
    OR ENDS_WITH(hunt_path, "/.continue/config.json")
    OR ENDS_WITH(hunt_path, "/.continuerc.json")
| EVAL hunt_group = CASE(ENDS_WITH(hunt_path, "/.claude/.credentials.json"), "credential", "configuration")
| LIMIT 10000

Elastic ES|QL 8.14+

AI file requests in Elastic Nginx logs, literal

Hunt both selected path groups in inbound access logs.

Requires: Nginx integration access streams; ECS url.path; read and ES|QL permission; no extra decoding.

FROM logs-nginx.access-*
| WHERE event.dataset == "nginx.access"
| EVAL hunt_original_target = url.original
| EVAL hunt_path = TO_LOWER(url.path)
| WHERE ENDS_WITH(hunt_path, "/.claude/.credentials.json")
    OR ENDS_WITH(hunt_path, "/.claude/settings.json")
    OR ENDS_WITH(hunt_path, "/.claude/settings.local.json")
    OR ENDS_WITH(hunt_path, "/.claude.json")
    OR ENDS_WITH(hunt_path, "/claude_desktop_config.json")
    OR ENDS_WITH(hunt_path, "/.mcp.json")
    OR ENDS_WITH(hunt_path, "/.cursor/mcp.json")
    OR ENDS_WITH(hunt_path, "/.continue/config.yaml")
    OR ENDS_WITH(hunt_path, "/.continue/config.json")
    OR ENDS_WITH(hunt_path, "/.continuerc.json")
| EVAL hunt_group = CASE(ENDS_WITH(hunt_path, "/.claude/.credentials.json"), "credential", "configuration")
| LIMIT 10000

CrowdStrike Falcon LogScale / Next-Gen SIEM CQL

AI file requests in CrowdStrike ingested Apache logs

Hunt both selected path groups in inbound access logs.

Requires: Third-party Apache Common/Combined Log Format logs with preserved @rawstring; select repository/view and verify #logtype; compatible CQL functions; search and ingestion entitlement.

#logtype="apache-access-log" @rawstring=*
| regex("^(?<hunt_client>[^ ]+) [^ ]+ [^ ]+ \\[[^\\]]+\\] \"(?<hunt_method>[^ ]+) (?<hunt_target>[^ \"]+) HTTP/[0-9.]+\"", field=@rawstring, flags="", strict=true)
| hunt_target=*
| regex("^(?:[^ ]+) [^ ]+ [^ ]+ \\[[^\\]]+\\] \"(?:[^ ]+) (?:[^ \"]+) HTTP/[0-9.]+\" (?<hunt_status>[0-9]{3})(?: |$)", field=@rawstring, flags="", strict=false)
| regex("^(?<hunt_raw_path>/[^?]*)", field=hunt_target, flags="", strict=true)
| hunt_raw_path=*
| urlDecode(field=hunt_raw_path, as=hunt_decoded_path)
| hunt_decoded_path=*
| lower(field=hunt_decoded_path, as=hunt_path, locale="en")
| case {
    text:endsWith(string=hunt_path, substring="/.claude/.credentials.json") | hunt_group := "credential";
    text:endsWith(string=hunt_path, substring="/.claude/settings.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.claude/settings.local.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.claude.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/claude_desktop_config.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.mcp.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.cursor/mcp.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.continue/config.yaml") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.continue/config.json") | hunt_group := "configuration";
    text:endsWith(string=hunt_path, substring="/.continuerc.json") | hunt_group := "configuration";
}

Python 3.10+ / local staging

Check a staged release for selected AI tool paths

Report selected filenames without opening artifact contents.

Requires: Explicit unpacked staging directory; Python; adjacent selected_paths.json.

YARA is not included: a text marker does not establish that an artifact contains a secret. The filename inventory reports candidates without opening their contents. The selected path list is a starting point to extend for your own tools.

Test the alert without putting a real key at risk

In an approved test environment, place an inert marker file in a monitored location and make an authorized request to it. Give the test an owner and a time window. Confirm that the responder receives the path, the response context, and an assigned next action. Use no production credential, and remove the test placement when the exercise is complete.

The point is to test the boundary and the handoff. A notification saying “someone asked for a file” should lead to a person who can establish whether the file was served and what to do about it.

ONE APPLICATION / ONE RELEASE / ONE NAMED OWNER

Make the check repeatable

Use the worksheet to record the release, private paths, public response, evidence location, and owner. Pair it with the standard rules and artifact checker.

Download the worksheet PDFOpen the web worksheetRead the carousel transcriptGet the rules and checks

Which step does your team check today: the repository, the built artifact, or the public response?

Sources and counting method

We independently recounted retained exports for September 8 through October 7, 2026, UTC. There were 2,441 in-window exported rows; collapsing identical full exported rows left 2,317. The retained AI-tooling filename classifier selected 1,429 rows, including 21 whose path exactly equaled /.claude/.credentials.json. Exported paths were limited to 120 characters. Deduplication is at the exported-row level, not a claim that every retained row represents a distinct person, original request, or source system. Response bodies are not available in this export, and status coverage is incomplete.

This is a refreshed analysis of a retained observation window, not a November traffic count. The graph uses the selected configuration-path dataset. Classifier details and a deidentified aggregate CSV accompany this field note; the raw operational records remain private. The deployment workflow, owner checklist, and test exercise are our recommended practice.

Primary documentation checked November 1, 2026: Claude Code credential storage, Cursor MCP configuration, Continue configuration formats, Docker build contexts, and OWASP secrets management.