Detections
Explore public detection logic contributed by the community across SIEM and rule languages.
2,982 detections
Filters
Last updated
All Time
Detection languages
1,486
513
403
239
167
Contributors
986
658
304
198
109
Categories
916
822
673
590
452
Platforms
39,934
7,265
6,535
3,866
3,645
2,982
Products / Services
1,229
703
443
309
200
MITRE Techniques
1,053
824
459
384
348
CVEs
17
16
16
10
10
IDS Classtypes
24
7
7
4
3
IDS Protocols
49
1
Detects attempts to perform prompt injection or meta-prompt probing against an LLM gateway to extract confidential system instructions, as well as instances where the LLM response echoes sensitive system-level boilerplate consistent with a successful prompt extraction attack.
Detects anomalous high-frequency requests to machine learning model endpoints from a single client, a pattern indicative of model extraction (stealing) attacks. This rule monitors inference request paths and success statuses to identify potential systematic input-space probing.
Detects adversarial attempts to poison Retrieval-Augmented Generation (RAG) knowledge bases. This includes the ingestion of documents containing prompt injection directives (e.g., 'ignore previous instructions'), hidden malicious markup intended to manipulate LLM inference, unauthenticated data ingestion, or abnormal volume spikes from a single source.
Detects Large Language Model (LLM) responses that contain sensitive information such as PII, internal proprietary markers, or credentials. This rule monitors model output for patterns matching social security numbers, credit card numbers, private keys, API keys, and internal company confidentiality labels, which could indicate data leakage or unintended disclosure.
Detects anomalous high-volume API requests from a single user/key to AI model inference endpoints. The rule tracks query diversity and total volume within an hour to identify behavior consistent with model extraction or cloning (AML.T0024.002) via systematic input/output pair collection.
Detects anomalous high-frequency queries to machine learning model inference endpoints involving probability or confidence score outputs. This behavior is indicative of a membership inference attack, where an adversary attempts to determine if specific data samples were included in a model's training set by harvesting confidence scores.
Detects high-frequency, sustained queries to an ML inference endpoint from a single identity, indicative of gradient-free training data reconstruction or model inversion attacks. The rule triggers when a specific user or API key performs over 100 successful requests within a 15-minute window, where the responses include granular confidence, probability, or logit values, which are typical targets for reconstructing training data.
Detects anomalous high-frequency queries to machine learning model inference endpoints involving probability or confidence score outputs. This behavior is indicative of a membership inference attack, where an adversary attempts to determine if specific data samples were included in a model's training set by harvesting confidence scores.
Detects high-frequency, sustained queries to an ML inference endpoint from a single identity, indicative of gradient-free training data reconstruction or model inversion attacks. The rule triggers when a specific user or API key performs over 100 successful requests within a 15-minute window, where the responses include granular confidence, probability, or logit values, which are typical targets for reconstructing training data.
Detects potential indirect prompt injection attacks against a Large Language Model (LLM) where untrusted context (e.g., RAG-retrieved documents, web content, or tool outputs) containing instruction-override or persona-switch keywords is ingested. This is followed by the LLM performing an anomalous privileged tool call within the same session, indicating a possible compromise of agent autonomy or unauthorized action execution.
Detects attempts to trigger a poisoned machine learning model through anomalous inference requests. The rule identifies inference API calls flagged as statistically anomalous by an upstream detector (e.g., unusual pixel patterns, token sequences, or steganographic payloads) that are correlated with a specific, spike-related output class across multiple distinct users.
Detects potential LLM jailbreak attempts by identifying a high frequency of guardrail-blocked responses followed by user prompts containing known jailbreak strategy fingerprints (such as roleplay, persona switching, or instruction overrides) within a 15-minute window.
Detects a sudden, anomalous spike in low-confidence actions or pending reviews originating from a single source within an AI agent action queue. This activity is indicative of a 'chaff' flooding attack, where an adversary submits a high volume of low-quality or irrelevant requests to overwhelm human-in-the-loop review systems, potentially causing a denial-of-service or masking malicious activity.
Detects anomalous usage patterns for AI-platform API keys consistent with account takeover, specifically identifying concurrent multi-geographic access and the exercise of elevated administrative, write, or model-registry scopes within a short timeframe.
This rule detects when a user grants OAuth application permissions for high-risk scopes (e.g., Mail, Files) to an application that was registered within the last 7 days. It correlates this event with subsequent suspicious actions performed by the same user, specifically mailbox rule modifications or file download activity, which are common indicators of post-compromise activity following OAuth consent phishing.
Detects MFA fatigue or push-bombing attacks where a user account experiences an abnormally high volume of MFA request denials in a short duration (10 minutes), followed by a successful authentication event. The rule specifically looks for evidence where the successful authentication potentially originates from a different IP address than the failed attempts.
Detects high-frequency, automated API requests to commercial Large Language Model (LLM) endpoints from a single source IP. The rule identifies suspicious activity characterized by a rapid rotation of multiple API tokens, often indicative of an intermediary proxy or middleware service attempting to bypass provider-specific usage limits, billing controls, or safety guardrails.
Detects instances where an AI agent invokes sensitive administrative or system tools (shell execution, file writing, network communication, or email sending) shortly after ingesting untrusted external content, in the absence of explicit, logged user instructions for those specific actions. This behavior is a primary indicator of indirect prompt injection attacks aimed at achieving command execution or C2-like behavior.
Detects high-volume, systematic querying of LLM inference endpoints, which may indicate unauthorized model extraction, distillation attempts, or targeted data exfiltration from an LLM.
Detects the granting of highly privileged OAuth or API scopes to AI-related applications (such as Copilot plugins, browser-automation agents, or AI connectors) immediately followed by high-volume data access activities (file reads or email transfers) performed by the same user account. This behavior is indicative of an attacker weaponizing a compromised identity to grant an external AI agent broad access to sensitive data for exfiltration or automated reconnaissance.
This rule detects potential prompt injection attacks against internal AI-powered assistants, agents, or LLM gateways. It monitors network HTTP requests for common patterns used to override system instructions, bypass safety filters, or force the model to adopt a privileged, unauthorized persona (e.g., administrator or debug mode). These techniques are often used in social engineering to manipulate the AI's output or security posture.
Page 47 of 142