Detections
Explore public detection logic contributed by the community across SIEM and rule languages.
2,588 detections
Filters
Last updated
All Time
Detection languages
1,194
459
310
204
161
Contributors
487
455
323
171
123
Categories
1,884
937
415
400
360
Platforms
1,349
560
343
329
257
Products / Services
601
314
306
286
271
MITRE Techniques
18,116
15,636
12,583
8,393
6,221
2,588
CVEs
9
5
5
5
2
IDS Classtypes
142
96
49
11
6
IDS Protocols
177
49
44
17
4
Detects a multi-stage, automated attack chain performed by AI-based development agents (e.g., Claude, Copilot, Cursor). The rule correlates process discovery, credential access, file collection (staging), and exfiltration over web services within a tight 15-minute window, identifying potential autonomous exploitation and data exfiltration. Covered T1046, T1552.001, T1078, T1005, T1567
Detects a multi-stage, automated attack chain performed by AI-based development agents (e.g., Claude, Copilot, Cursor). The rule correlates process discovery, credential access, file collection (staging), and exfiltration over web services within a tight 15-minute window, identifying potential autonomous exploitation and data exfiltration. Covered T1046, T1552.001, T1078, T1005, T1567
Detects a multi-stage, automated attack chain performed by AI-based development agents (e.g., Claude, Copilot, Cursor). The rule correlates process discovery, credential access, file collection (staging), and exfiltration over web services within a tight 15-minute window, identifying potential autonomous exploitation and data exfiltration. Covered T1046, T1552.001, T1078, T1005, T1567
Detects a multi-stage, automated attack chain performed by AI-based development agents (e.g., Claude, Copilot, Cursor). The rule correlates process discovery, credential access, file collection (staging), and exfiltration over web services within a tight 15-minute window, identifying potential autonomous exploitation and data exfiltration. Covered T1046, T1552.001, T1078, T1005, T1567
Detects instances where a single process on a device communicates with two or more distinct hosted LLM provider APIs (OpenAI, Anthropic, DeepSeek, Google) within a 24-hour window. This behavior is indicative of multi-model orchestration or 'consensus' techniques, which may be employed by sophisticated implants for automated tasking, data processing, or evasion.
Detects instances where AI-based coding or agent-driven tools (e.g., Cursor, Claude Code, AutoGPT) perform suspicious file archiving/packaging activities followed by network egress to unauthorized or known exfiltration-prone services. This helps identify potential data exfiltration by AI agents or compromised development environments.
Detects instances where AI-based coding or agent-driven tools (e.g., Cursor, Claude Code, AutoGPT) perform suspicious file archiving/packaging activities followed by network egress to unauthorized or known exfiltration-prone services. This helps identify potential data exfiltration by AI agents or compromised development environments.
Detects instances where AI coding and agent tools (e.g., Claude Code, Cursor, Copilot CLI, Ollama, LangGraph) create archive or database files containing potentially sensitive data and subsequently initiate network connections to non-AI-provider or non-developer services (exfiltration services or non-standard port connections).
Detects DNS queries, TLS SNI headers, and HTTP host headers associated with the known exfiltration domain 't.m-kosche.com' linked to the Mini Shai-Hulud malware.
Detects potential embedding inversion or extraction attempts by identifying callers exhibiting abnormally high volumes of similarity or vector search queries against RAG-enabled vector stores or cognitive search APIs within a short timeframe.
Detects potential embedding inversion or extraction attempts by identifying callers exhibiting abnormally high volumes of similarity or vector search queries against RAG-enabled vector stores or cognitive search APIs within a short timeframe.
The following analytic detects potential data exfiltration using plain HTTP POST requests. It leverages network traffic logs, specifically monitoring the `stream_http` data source for POST methods containing suspicious form data such as "wermgr.exe" or "svchost.exe". This activity is significant because it is commonly associated with malware like Trickbot, trojans, keyloggers, or APT adversaries, which use plain text HTTP POST requests to communicate with remote C2 servers. If confirmed malicious, this activity could lead to unauthorized data exfiltration, compromising sensitive information and potentially leading to further network infiltration.
This rule monitors for file creation or existence patterns in the user's Local Temp directory that are commonly associated with infostealer malware, specifically looking for naming conventions like 'evo_cl_*.exe', 'evo_db_*.db', or the presence of a 'stolen.zip' file within subdirectories.
Detects anomalous, unauthorized access to machine learning model artifacts (e.g., .onnx, .pt, .h5) within object storage or model registries. The rule identifies suspicious patterns including high-volume file reads by non-ML team identities, lack of multi-factor authentication, or access from geolocations outside approved operating regions, indicative of potential model exfiltration or intellectual property theft.
Detects newly consented OAuth applications that immediately perform mass access or exfiltration of mail and file data across multiple mailboxes or drives within a short period, potentially indicating an attacker utilizing a malicious OAuth application.
This rule establishes a daily baseline of request volumes to mail.google.com per user and per host using proxy logs. It identifies anomalous spikes in traffic or unexpected users/hosts accessing personal webmail, which may indicate data exfiltration using a legitimate web service.
Detects anomalous, high-volume file access by a single process, followed by local file compression or archiving, and subsequent outbound data transfer to a non-internal IP address. This pattern is characteristic of automated agents (like those used in the GTG-1002/Anthropic AI-orchestrated campaigns) performing rapid data triage and staging for exfiltration.
This rule detects outgoing HTTP POST requests that contain a ZIP file with a filename prefix of 'StolenData_' in the request body. This is a known behavior of the TokenGrabber stealer family, which exfiltrates collected victim data (e.g., browser credentials, session tokens) to a remote server, often via a Discord webhook or similar infrastructure.
Detects anomalous process activity that reads sensitive developer secrets (such as Anthropic, OpenAI, Gemini, or Hugging Face API keys) from common configuration or environment files, followed by an outbound network connection to a non-private IP address. This behavior is consistent with credential harvesting and exfiltration tactics used in supply chain attacks to hijack AI services.
Detects anomalous, high-volume file access by a single process, followed by local file compression or archiving, and subsequent outbound data transfer to a non-internal IP address. This pattern is characteristic of automated agents (like those used in the GTG-1002/Anthropic AI-orchestrated campaigns) performing rapid data triage and staging for exfiltration.
Detects instances where a non-browser process initiates connections to commercial LLM provider APIs (DeepSeek, OpenRouter, Mistral) followed by a connection to Discord CDN/Webhook endpoints within a 15-minute window. This behavior is indicative of a process acting as an autonomous C2 agent that exfiltrates data after receiving instructions or processing information via an LLM.
Page 17 of 124