Detections
Explore public detection logic contributed by the community across SIEM and rule languages.
29 detections
Filters
Last updated
All Time
Detection languages
9
8
7
5
Contributors
14
9
4
1
1
Categories
16
11
10
6
4
Platforms
16
12
3
2
1
Products / Services
15
11
4
3
1
MITRE Techniques
18,116
15,638
12,585
8,395
6,225
Detects agentic AI jailbreak attempts that try to establish persistent control over M365 Copilot through rule injection, universal triggers, response automation, system overrides, and persona establishment techniques. The detection analyzes the PromptText field for keywords like "from now on," "always respond," "ignore previous," "new rule," "override," and role-playing commands (e.g., "act as," "you are now") that attempt to inject persistent instructions. The search computes risk by counting distinct jailbreak indicators per user session, flagging coordinated manipulation attempts.
Detects M365 Copilot jailbreak attempts through prompt injection techniques including rule manipulation, system bypass commands, and AI impersonation requests that attempt to circumvent built-in safety controls. The detection searches exported eDiscovery prompt logs for jailbreak keywords like "pretend you are," "act as," "rules=," "ignore," "bypass," and "override" in the Subject_Title field, assigning severity scores based on the manipulation type (score of 4 for amoral impersonation or explicit rule injection, score of 3 for entity roleplay or bypass commands). Prompts with a jailbreak score of 2 or higher are flagged, prioritizing the most severe attempts to override AI safety mechanisms through direct instruction injection or unauthorized persona adoption.
Detects M365 Copilot impersonation and roleplay jailbreak attempts where users try to manipulate the AI into adopting alternate personas, behaving as unrestricted entities, or impersonating malicious AI systems to bypass safety controls. The detection searches exported eDiscovery prompt logs for roleplay keywords like "pretend you are," "act as," "you are now," "amoral," and "roleplay as" in the Subject_Title field. Prompts are categorized into specific impersonation types (AI_Impersonation, Malicious_AI_Persona, Unrestricted_AI_Persona, etc.) to identify attempts to override the AI's safety guardrails through persona injection attacks.
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 web requests targeting common LLM/AI inference API endpoints containing strings indicative of prompt injection attempts, such as attempts to bypass filters, exfiltrate system instructions, or force model behavior changes (e.g., DAN mode).
Detects the presence of zero-width, non-printing, or bidirectional-override Unicode characters in command lines, commonly used for obfuscation or prompt injection.
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 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 invisible/non-printing Unicode codepoints (zero-width chars, bidi overrides, variation selectors, Unicode tag characters, and Private Use Area icon-font codepoints) embedded in process command lines - covering both classic command/script obfuscation and the emerging use of hidden codepoints to smuggle AI prompt-injection payloads into content later processed by AI copilots/agents. Tightened to require 2+ high-confidence codepoint hits (or 1 from a known scripting host/LOLBin), and gates Private-Use-Area-only matches (a common legitimate icon-font false-positive source) to scripting hosts with 5+ hits, maps to T1027.018.
Detects the registration or installation of an MCP (Model Context Protocol) server or agent tool containing suspicious prompt-injection strings (e.g., instructions to ignore safety guardrails or role overrides), immediately followed by the execution of shell commands or unauthorized network connections originating from that tool.
Detects a potential AI agent data poisoning event where a file is downloaded from an untrusted source (browser or sync client) and subsequently read within one hour by a process associated with AI agent activity (e.g., semantic_kernel, SKAgent). This pattern is often used to inject malicious context or prompt injection vectors into RAG (Retrieval-Augmented Generation) systems.
Detects the presence of zero-width, non-printing, or bidirectional-override Unicode characters in command lines, commonly used for obfuscation or prompt injection.
Detects invisible/non-printing Unicode codepoints (zero-width chars, bidi overrides, variation selectors, Unicode tag characters, and Private Use Area icon-font codepoints) embedded in process command lines - covering both classic command/script obfuscation and the emerging use of hidden codepoints to smuggle AI prompt-injection payloads into content later processed by AI copilots/agents. Tightened to require 2+ high-confidence codepoint hits (or 1 from a known scripting host/LOLBin), and gates Private-Use-Area-only matches (a common legitimate icon-font false-positive source) to scripting hosts with 5+ hits, maps to T1027.018.
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.
Detects invisible/non-printing Unicode codepoints (zero-width chars, bidi overrides, variation selectors, Unicode tag characters, and Private Use Area icon-font codepoints) embedded in process command lines - covering both classic command/script obfuscation and the emerging use of hidden codepoints to smuggle AI prompt-injection payloads into content later processed by AI copilots/agents. Tightened to require 2+ high-confidence codepoint hits (or 1 from a known scripting host/LOLBin), and gates Private-Use-Area-only matches (a common legitimate icon-font false-positive source) to scripting hosts with 5+ hits, maps to T1027.018.
Detects invisible/non-printing Unicode codepoints (zero-width chars, bidi overrides, variation selectors, Unicode tag characters, and Private Use Area icon-font codepoints) embedded in process command lines - covering both classic command/script obfuscation and the emerging use of hidden codepoints to smuggle AI prompt-injection payloads into content later processed by AI copilots/agents. Tightened to require 2+ high-confidence codepoint hits (or 1 from a known scripting host/LOLBin), and gates Private-Use-Area-only matches (a common legitimate icon-font false-positive source) to scripting hosts with 5+ hits, maps to T1027.018.
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.
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.
Detects the creation or modification of documents in common formats (e.g., .txt, .md, .docx, .pdf) that contain strings indicative of prompt-injection attempts. These strings are commonly used to manipulate the behavior of AI agents or Large Language Models (LLMs) that may process or index these files automatically.
Page 1 of 2