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
10,660
9,758
6,692
4,418
3,748
1,084
MITRE Techniques
278
229
180
177
151
CVEs
17
15
13
13
8
IDS Classtypes
4
IDS Protocols
4
Detects anomalous high-volume API requests to machine learning model inference endpoints (e.g., predict, classify, score). This behavior is characteristic of adversarial example crafting, where an attacker sends rapid, near-duplicate requests to estimate gradients or probe confidence scores to manipulate model output.
This rule detects unauthorized or anomalous file modifications or creations to machine learning training datasets (e.g., fine-tuning data, training corpora) by users not belonging to the authorized data engineering group, occurring within one hour of a scheduled training or fine-tuning job execution. This behavioral pattern is indicative of potential data poisoning, where an adversary attempts to inject malicious data into the training set to bias or compromise the resulting model.
Detects malicious instruction-override payloads, hidden HTML comments, or obscured text (e.g., zero-font size) in third-party content ingested into LLM/RAG systems, commonly used for indirect prompt injection.
Detects potential ML model-stealing attacks against inference API endpoints. The rule identifies anomalous behavior characterized by an extremely high volume of diverse queries from a single API key over a 24-hour period. Such patterns are consistent with systematic decision-boundary probing or input-space coverage attempts used for training a replica model (distillation or knockoff attacks) without direct access to model weights.
This rule detects modifications to critical AI agent configurations—such as system prompts, tool allowlists, permission scopes, or model parameters—by users who are not part of the authorized 'agent-ops' group. It flags potentially high-risk changes, particularly when those changes fall outside established maintenance windows or involve an expansion of agent permissions.
Detects attempts to bypass LLM safety guardrails and system instructions via malicious user prompts containing injection phrases or privilege escalation keywords, which could lead to unauthorized tool execution or sensitive data disclosure.
This rule detects unauthorized or malicious agentic AI behavior where prompt injection or poisoned content causes an AI agent to execute excessive tool calls or enter an infinite loop of tool invocations. This behavior can lead to compute resource exhaustion, increased operational costs, and potential data exfiltration via excessive querying.
Detects adversarial AI cost harvesting (AML.T0034.000) by identifying anomalous volumes of low-complexity queries against an AI inference API that result in a significant, unexplained spike in billed costs. This behavior mimics denial-of-service tactics aimed at exhausting compute capacity and financial resources.
Detects ingestion of documents containing potential LLM prompt injection payloads (using zero-width characters or HTML comments to hide imperative instructions) into a RAG system, followed by anomalous LLM behavior such as unauthorized tool execution or unintended data disclosure.
Detects ingestion of documents containing potential LLM prompt injection payloads (using zero-width characters or HTML comments to hide imperative instructions) into a RAG system, followed by anomalous LLM behavior such as unauthorized tool execution or unintended data disclosure.
Detects ingestion of documents containing potential LLM prompt injection payloads (using zero-width characters or HTML comments to hide imperative instructions) into a RAG system, followed by anomalous LLM behavior such as unauthorized tool execution or unintended data disclosure.
Detects anomalous AI model inference activity where a caller submits a high volume of near-duplicate queries while frequently requesting raw confidence scores. This behavior is indicative of a membership-inference privacy attack, where an adversary attempts to determine if specific data points were included in the model's training set by analyzing query response patterns.
Detects LLM jailbreak attempts using a 'crescendo' multi-turn strategy. The rule monitors for sequences where an initial prompt contains known jailbreak-intent keywords leading to an AI refusal, followed shortly by a compliant AI response to a rephrased request within the same conversation session.
Detects adversaries attempting to override LLM system instructions by submitting crafted prompts containing known injection patterns (e.g., 'ignore previous instructions', role-override). The rule tracks anomalous interaction patterns identified by multiple occurrences or a variety of override phrases within a short timeframe.
Detects the ingestion of documents into production RAG (Retrieval-Augmented Generation) indexes from untrusted or external-facing sources, which may indicate an attempt to perform RAG poisoning attacks.
This rule detects potential cost harvesting or service abuse against AI gateway endpoints. It establishes a baseline of typical user activity over a 30-day period and flags anomalies where current request volume (in the last 1 hour or 24 hours) significantly exceeds the user's established P95 baseline. To reduce noise, it specifically identifies scenarios where these high-frequency requests are associated with low-complexity, short prompts, which are indicative of automated probing or resource-intensive scraping rather than legitimate high-utility usage.
This rule detects anomalous behavior in AI agent tool invocations, specifically monitoring for high volumes of tool calls, high costs associated with token consumption, or suspicious tool interactions following a flagged indirect prompt injection. This behavior is indicative of potential resource exhaustion (denial of service) or abusive agent behavior as defined by the MITRE ATLAS framework for AI security.
Detects suspicious AI agent behavior by identifying two primary scenarios: 1) The invocation of sensitive tools (e.g., code execution, file deletion) within a short timeframe following a flagged prompt injection or jailbreak attempt in the same conversation; or 2) The invocation of a tool by an AI agent that is not explicitly defined in its allowed tool list for the given workflow.
This rule detects potential Denial of Service (DoS) attacks against AI inference endpoints. It monitors for sustained high latency and high error rates combined with a high volume of 'heavy' requests (those exceeding a token/character count threshold) from specific source IPs. This pattern suggests an attempt to exhaust system resources by forcing the processing of complex or overly large prompts.
This rule detects potential Denial of Service (DoS) attacks against AI inference endpoints. It monitors for sustained high latency and high error rates combined with a high volume of 'heavy' requests (those exceeding a token/character count threshold) from specific source IPs. This pattern suggests an attempt to exhaust system resources by forcing the processing of complex or overly large prompts.
This rule monitors package installation logs for indicators of packages potentially generated or suggested by AI tools (e.g., Copilot, code assistants) being installed in a target environment. It specifically looks for a low volume of installations (<=3) for packages that have been published within the last 14 days, which is a pattern often associated with the 'Publish Hallucinated Entities' technique in AI systems, where malicious or hallucinated code packages are introduced into the supply chain.
Page 7 of 52