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
4,418
3,749
1,084
MITRE Techniques
278
229
180
177
151
CVEs
17
15
13
13
8
IDS Classtypes
4
IDS Protocols
4
Detects anomalous AI agent behavior where the agent ingests untrusted external content (e.g., tool outputs, RAG retrievals, or web fetches) and subsequently invokes high-risk or sensitive capabilities (such as shell execution, secret reading, or unauthorized network calls) within the same session. This pattern is indicative of an indirect prompt injection attack attempting to leverage the agent's internal tool-calling mechanisms for unauthorized actions.
Detects anomalous AI agent behavior where the agent ingests untrusted external content (e.g., tool outputs, RAG retrievals, or web fetches) and subsequently invokes high-risk or sensitive capabilities (such as shell execution, secret reading, or unauthorized network calls) within the same session. This pattern is indicative of an indirect prompt injection attack attempting to leverage the agent's internal tool-calling mechanisms for unauthorized actions.
Detects unauthorized or suspicious modifications to AI model artifacts, including fine-tuned weights and LoRA adapters. This rule identifies tampering attempts via unsigned uploads, bypassing CI/CD production deployment approvals, or mismatches in hash/checksum compared to known-good versions.
Detects an AI agent identity accessing an unusually high number of distinct internal repositories, such as wikis, ticketing systems, code repositories, or customer data stores within a 15-minute window. This behavior is consistent with automated discovery or collection (T1213) by an AI agent.
Detects the presence of a specific, hardcoded Ed25519 public key associated with malicious cross-platform payloads (npm, Go modules, Terraform providers) that use Slack for C2 communication.
Detects AI agents attempting to disable or bypass security guardrails, approval gates, or content filters. Additionally, it identifies high-risk or destructive actions taken by an AI agent that lack the required preceding human-approval audit record.
Detects anomalous behavior by an AI agent identity by monitoring for off-hours activity, significant statistical spikes in task invocation volumes compared to a 30-day baseline, or the execution of task categories not previously associated with that specific agent. Such behaviors may indicate account hijacking or unauthorized command injection.
Detects attempts to bypass AI agent safety guardrails by using obfuscated payloads (Base64/hex) or known jailbreak/roleplay trigger phrases. The rule specifically monitors sequences where an initial prompt is blocked by safety filters and is immediately followed by a structurally similar, potentially re-encoded prompt that is subsequently allowed, indicating an active filter evasion attempt.
Detects potential poisoning of an AI agent's RAG knowledge base or vector store through suspicious bulk ingestion or injection of imperative instruction-style content. The rule correlates these ingestion events with subsequent anomalous tool usage patterns by the same agent session, indicating a possible prompt injection or data manipulation attack aimed at altering agent behavior.
Detects malicious JavaScript injection on e-commerce checkout or payment pages designed to harvest sensitive payment card information (PAN, CVV, expiry) and exfiltrate the stolen data to an external endpoint. This behavior is indicative of a digital credit card skimmer (e.g., Magecart style) deployed following a compromise of the web application.
Detects anomalous AI inference API activity characterized by high request volume and high input diversity (querying a wide range of input space). This behavior is indicative of a model replication or extraction sweep aimed at exfiltrating model characteristics or intellectual property.
Detects potential AI model backdoor trigger attempts by identifying inference requests from a single caller that result in high-confidence, rare-label predictions across multiple distinct input artifacts. This pattern is indicative of an adversary probing or activating a poisoned model by injecting specific triggers designed to force anomalous outputs.
Detects unauthorized or unreviewed modifications to AI model weights, retraining jobs, or pushes to the model registry within MLOps pipelines. This activity is monitored to prevent AI model weight poisoning, ensuring that all model-related changes undergo required peer-review or approval processes before deployment.
Detects adversarial evasion attempts against AI-based classification models by identifying inputs that result in a benign or low-risk prediction with confidence scores just below a configured threshold. The detection specifically looks for artifacts that share a hash prefix with previously confirmed malicious samples, indicating a potential iterative perturbation technique used to bypass model boundaries.
Detects unauthorized modifications or non-pipeline uploads to an ML model registry. The rule flags events where model artifacts lack valid signatures, fail checksum verification, or are uploaded/promoted by non-authorized identities or without required CI/CD provenance, indicating potential ML model backdooring or poisoning.
Detects potential data exfiltration from ML inference APIs by identifying responses to untrusted clients containing patterns indicative of memorized PII (SSN, Email, Credit Card) or detecting automated, repeated extraction-style queries.
Detects attempts to bypass LLM safety alignment by utilizing roleplay or hypothetical framing (e.g., DAN mode) alongside evidence of prompt iteration, identified by repeated moderation blocks within the same session.
Detects LLM input containing known instruction-override patterns (e.g., 'ignore previous instructions', 'DAN' jailbreak) paired with requests for privileged system operations, indicating a malicious attempt to bypass guardrails and gain unauthorized control over the LLM model.
Detects LLM input containing known instruction-override patterns (e.g., 'ignore previous instructions', 'DAN' jailbreak) paired with requests for privileged system operations, indicating a malicious attempt to bypass guardrails and gain unauthorized control over the LLM model.
This rule detects the unauthorized download of complete ML model checkpoint or weight files from a model registry or object storage. It monitors for common model file extensions in download events that are not associated with a legitimate deployment justification or a pre-approved change management window, helping to identify potential intellectual property theft or model exfiltration attempts.
Detects anomalous access patterns to machine learning model inference endpoints, including unauthorized access attempts such as credential failures, scope violations, and quota exhaustion, indicating potential misuse or compromise of private model APIs.
Page 9 of 52