Detections
Explore public detection logic contributed by the community across SIEM and rule languages.
1,084 detections
Filters
Last updated
All Time
Detection languages
418
193
163
156
81
Contributors
381
251
92
71
40
Categories
498
271
255
135
123
Platforms
443
441
267
121
114
Products / Services
10,665
9,758
6,692
4,421
3,749
1,084
MITRE Techniques
278
229
180
177
151
CVEs
17
15
13
13
8
IDS Classtypes
4
IDS Protocols
4
Detects the insertion of malicious prompt injection payloads—containing imperative commands to AI agents or hidden unicode/non-printable characters—into wiki pages, ticketing systems, or other shared knowledge bases. These payloads are designed to be ingested by AI agents or Retrieval-Augmented Generation (RAG) pipelines, leading to unauthorized instructions or behavioral overrides upon retrieval.
Detects credential-spraying activity against WordPress sites where an actor performs a sequence of POST requests targeting both xmlrpc.php (typically the 'wp.getUsersBlogs' method) and wp-login.php. The rule specifically identifies attackers that attempt to evade detection or increase success rates by falling back to wp-login.php after initially using XML-RPC within a short duration.
This rule detects a potential vishing attack by correlating suspicious IT helpdesk requests (e.g., password or MFA reset without manager approval or verified callback) with subsequent sensitive Entra ID identity changes (e.g., self-service password resets, MFA device registration, or new device registration) occurring for the same user within a short time window. This activity is indicative of an attacker attempting to social-engineer helpdesk staff to bypass authentication controls and compromise a user account.
Detects critical log entries on MikroTik RouterOS devices that contain 'flagged' status messages following a system upgrade or maintenance. This pattern is indicative of potential post-exploitation activity or post-update integrity warnings identified in recent vulnerabilities affecting RouterOS.
This rule detects scenarios where an AI agent executes shell commands or system-level tools immediately after receiving a prompt injection attack from an untrusted source. It correlates agent audit logs—specifically identifying tool calls flagged for prompt injection—with actual endpoint process creation events. It flags process execution as 'High' severity if the command line matches common malicious patterns like encoded commands, IEX, or attempts to access credential material (e.g., lsass, mimikatz).
This rule detects scenarios where an AI agent executes shell commands or system-level tools immediately after receiving a prompt injection attack from an untrusted source. It correlates agent audit logs—specifically identifying tool calls flagged for prompt injection—with actual endpoint process creation events. It flags process execution as 'High' severity if the command line matches common malicious patterns like encoded commands, IEX, or attempts to access credential material (e.g., lsass, mimikatz).
Detects the presence of a shared client-side JavaScript framework used in 'ClickFix' MaaS (Malware-as-a-Service) panels, specifically associated with the Exvicy and ErrTraffic campaigns. The rule identifies specific function patterns (e.g., xdReq, xdDec, waitForToken) and state machine logic (tokens, deduplication hashing) commonly used for token interception and clipboard interaction.
Detects unauthenticated HTTP POST requests targeting SharePoint WebPartPage-derived endpoints using the 'webPartMarkup' parameter. This activity attempts to exploit a vulnerability (CVE-2026-65660) where anonymous access to specific pages bypasses authentication, allowing an attacker to invoke internal methods for malicious register-directive injection.
Detects anomalous, high-volume, low-severity events flagged for human review originating from a single AI agent or source within a 10-minute window. This behavior is indicative of an adversary flooding the human-in-the-loop (HITL) review queue with chaff data (noise) to exhaust analyst capacity or conceal malicious activity.
This rule detects potential AI model training data poisoning attempts by monitoring data pipeline audit logs for suspicious activities. Specifically, it flags bulk modifications to production training sets, ingestion from newly introduced external sources, or failures in data integrity checksums, which are indicative of attempts to influence or corrupt machine learning model training pipelines.
This rule detects potential Training Data Poisoning (AML.T0020) by monitoring for anomalous activities in ML data pipelines. It triggers when ingestion jobs utilize untrusted sources, unregistered source paths, unauthorized accounts write to datasets, or when significant data volume bursts or schema/checksum drifts occur immediately prior to a scheduled training run.
Detects instances where an LLM agent initiates sensitive or out-of-sandbox tool calls (such as file system access or arbitrary command execution) following input that contains suspicious directive-override or prompt injection patterns. This behavior indicates that the agent has been compromised by a prompt injection attack, allowing an unauthorized actor to control the agent's tool-calling capabilities.
Detects Cost Harvesting activity (AML.T0034) targeting ML inference services, characterized by a significant spike in request volume or compute consumption coupled with low payload uniqueness. This behavior suggests an adversary is intentionally driving up billing costs by submitting redundant or low-value requests to the API.
Detects potential Denial of ML Service (DoMLs) attacks targeting inference endpoints. The rule identifies anomalous behavior characterized by large request payloads (oversized requests) correlated with increased latency or high GPU utilization, suggesting an attempt to exhaust compute resources and impact model availability.
Detects Large Language Model (LLM) responses that contain instructions designed to propagate the current prompt to other users, systems, or agents. This pattern is indicative of a self-replicating prompt worm, which aims to spread malicious instructions across an AI ecosystem by leveraging the model's ability to generate text that instructs subsequent interactions.
Detects instances where an AI agent or LLM invokes a high-privilege or sensitive tool (such as code execution, outbound HTTP, or credential store access) immediately following a prompt that exhibits characteristics of prompt injection or jailbreaking. This behavior matches MITRE ATLAS pattern AML.T0053, suggesting an attempt to use the AI agent as a pivot or execution vehicle for unauthorized actions.
This rule detects potential cost harvesting or unauthorized resource abuse by monitoring API endpoints (typically AI model gateways) for high-volume or resource-intensive queries associated with a specific identity. It correlates these patterns with billing or autoscaling trigger events (e.g., cost spikes, quota exhaustion) and identifies agentic sessions that exhibit abnormally long reasoning loops, which suggests an attempt to waste computational budget.
Detects recovered scripts/payloads that harvest environment variables, API keys, bearer tokens, and Kubernetes secrets/configmaps into a variable named LOOT
Detects instances where an AI agent's persistent memory store is updated with suspicious content (e.g., prompt injection, role overrides, or embedded credentials) by a user other than the authorized agent owner, followed by subsequent tool invocations in an independent session. This behavior aligns with prompt injection attacks aimed at manipulating agent logic and unauthorized resource access via the AI agent.
Detects malicious prompt injection attempts within an LLM conversation session. The rule monitors for ingested content containing common instruction-override keywords (e.g., 'ignore previous instructions', 'system prompt override'), followed immediately by the agent process executing unauthorized actions such as process launch, network connections, or file creation within the same session/user context.
Detects bulk insertions or content updates into vector databases, document stores, or knowledge bases that contain suspected instruction-override, jailbreak, or role-hijacking phrases, particularly when performed by unauthenticated or unknown identities.
Page 14 of 52