Skip to navigation

Threat reportThreat IntelligenceTL-2026-1635

"Living Off the AI": Attackers Abuse AI Agents, MCP Connectors, and Vibe-Coding Platforms for Offensive Ops

mediumACTIVE

Security researchers document an emerging attacker tradecraft pattern -- "living off the AI" -- in which adversaries piggyback on sanctioned AI assistant, agent, and Model Context Protocol (MCP) workflows instead of deploying custom malware. Techniques include prompt-born tool abuse, MCP/agent tooljacking via over-permissioned tool chains, memory/retrieval poisoning of vector stores, and AI-driven "vibe coding" platforms used to build phishing infrastructure and a Chrome credential-stealer proof-of-concept.

Key facts

IdentifiersTL-2026-1635
Also known asLiving Off the AI, LotAI, AI-native Living-off-the-Land
SeverityMEDIUM
Affected productsAnthropic Model Context Protocol SDKs (Python, TypeScript, Java, Rust); Wix / Base44 Base44 vibe-coding platform; Anysphere Cursor (AI coding IDE, MCP client); Codeium Windsurf (AI coding IDE, MCP client); Multiple LangChain, LiteLLM, IBM LangFlow (MCP-integrating frameworks)
Weaknesses (CWE)CWE-287, CWE-306, CWE-20, CWE-829
MalwareChrome credential stealer PoC (vibe-coded)
ToolsBase44, Cursor, IBM LangFlow, LangChain, LiteLLM, Lovable, Model Context Protocol (MCP), Netlify, Vercel, Windsurf
MITRE ATT&CK techniquesT1027 Obfuscated Files or Information, T1041 Exfiltration Over C2 Channel, T1068 Exploitation for Privilege Escalation, T1070 Indicator Removal, T1071.001 Web Protocols, T1082 System Information Discovery, T1083 File and Directory Discovery, T1102 Web Service, T1119 Automated Collection, T1140 Deobfuscate/Decode Files or Information, T1189 Drive-by Compromise, T1190 Exploit Public-Facing Application and 15 more in the attack flow
MotivationUNKNOWN
Target sectorstechnology, finance, enterprise-saas, government administration, health
Target regionsGlobal
Detection rules9 (1 shown free)
Sources10 cited sources

Detection coverage

As of 2026-02-06, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1635 across Splunk SPL, Microsoft KQL and Sigma, covering 25 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: MCP Tool-Response Prompt Injection / Tool-Poisoning Payload Ingestion

Detects MCP client/agent-runner log events where a tool response body contains imperative natural-language directives (e.g. 'agent, please...', 'assistant, now...', 'ignore previous instructions') instead of structured data — the hallmark of MCP Tool Poisoning / prompt-born tool abuse in which a compromised or malicious MCP server smuggles hidden instructions into the LLM's trusted context.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1566.002, T1199, T1203
  • Data sources: agent_platform, mcp_tool_logs
  • Severity: high
  • Confidence: medium

Splunk SPL

index=agent_platform sourcetype="mcp:tool_response" 
| regex response_body="(?i)(agent,?\s+(please|now)|assistant,?\s+(please|now)|ignore\s+(the\s+)?previous\s+instructions|disregard\s+prior)" 
| eval risk_reason="MCP tool response contains embedded imperative directive text" 
| table _time, mcp_server, mcp_server_id, tool_name, response_body, client_app, host, risk_reason 
| sort -_time

Microsoft KQL

MCPToolEvents
| where TimeGenerated > ago(24h)
| where EventType == "tool_response"
| where ResponseBody matches regex @"(?i)(agent,?\s+(please|now)|assistant,?\s+(please|now)|ignore\s+(the\s+)?previous\s+instructions|disregard\s+prior)"
| project TimeGenerated, McpServerId, ToolName, ResponseBody, ClientApp, DeviceName

Sigma

title: MCP Tool-Response Prompt Injection / Tool-Poisoning Payload
id: 8f3c1a2e-6b4d-4e9a-9c2f-1a7d3e5b9c01
status: experimental
description: Detects MCP tool responses carrying embedded natural-language imperative instructions rather than structured data, consistent with MCP Tool Poisoning / prompt-born tool abuse
references:
    - https://owasp.org/www-community/attacks/MCP_Tool_Poisoning
    - https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks
author: Threadlinqs Team
date: 2026-07-22
tags:
    - attack.initial-access
    - attack.t1566.002
    - attack.t1199
    - attack.tool.mcp
    - cve.na
logsource:
    category: application
    product: agent_platform
detection:
    selection_directive:
        EventType: 'tool_response'
        ResponseBody|re: '(?i)(agent,?\s+(please|now)|assistant,?\s+(please|now)|ignore\s+(the\s+)?previous\s+instructions|disregard\s+prior)'
    condition: selection_directive
falsepositives:
    - Legitimate tool documentation or help text that happens to use imperative phrasing
level: high

Known false positives:

  • Legitimate MCP tool descriptions or help text that use imperative phrasing without malicious intent

Tuning notes: Start with a broad regex on known imperative markers, then narrow to phrases observed in your MCP server allowlist violations; exclude servers on a vetted/pinned-hash registry to cut noise.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Over-Permissioned Agent Tool-Chain Fan-Out Following Untrusted Document Ingestion · unlock with TL Creds
  • Unvalidated-Trust-Score Vector-Store Write Preceding Agent Behavior Change (Memory/RAG Poisoning) · unlock with TL Creds
  • Vibe-Coding Platform Authentication Bypass via app_id-Only Registration/OTP Endpoint Abuse · unlock with TL Creds
  • Zero-Click Prompt-Injection-to-RCE Process Spawn from AI Coding IDE / MCP Client · unlock with TL Creds
  • Vibe-Coding-Generated Chrome Credential-Stealer Artifact Deployment · unlock with TL Creds
  • MCP Server Cloud-Account / Web-Service Infrastructure Stand-Up Preceding C2-Style Agent Traffic · unlock with TL Creds
  • Agent-Runner Registry Persistence Following Tool-Poisoned Context · unlock with TL Creds
  • Agent-Mediated SaaS-Routed Exfiltration (Cloud Camouflage) Following Poisoned Memory Reference · unlock with TL Creds

Indicators of compromise

As of 2026-02-06, Threadlinqs Intelligence tracks 25 indicators of compromise for TL-2026-1635.

25 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

In a February 2026 analysis, Etay Maor (VP of Threat Intelligence, Cato Networks) frames "living off the AI" as the AI-era successor to living-off-the-land (LotL) tradecraft: rather than deploying custom malware, attackers abuse the same sanctioned AI assistants, agents, and Model Context Protocol (MCP) connectors that organizations have deployed for productivity. Four converging technique classes are documented. First, prompt-born tool abuse: hidden instructions embedded in documents or webpages an agent ingests cause it to invoke connected tools -- querying internal RAG indexes for secrets, scheduling tasks, or exfiltrating data -- without tripping traditional EDR, because the actions are performed by an already-trusted agent identity. Second, MCP and agent "tooljacking": over-permissioned tool chains (unrestricted filesystem mounts, full-tenant search, unbounded web fetch) let an attacker nudge an agent into chaining tools in ways never intended by the integration's designer. This is closely related to the independently-documented "MCP Tool Poisoning" primitive, in which a malicious or compromised MCP server returns tool responses containing hidden natural-language instructions that land directly in the LLM's context and are treated as trusted input -- exploiting a structural trust gap between one-time, connect-time tool-description review and unvalidated, per-call tool-response content. Third, memory and retrieval poisoning: malicious "facts" seeded into shared vector stores or prior chat histories persist and reshape future agent behavior, suppressing warnings or inserting attacker-controlled exfiltration endpoints; this class has since been formalized industry-wide as OWASP Agentic AI Top 10 category ASI06 (Memory and Context Poisoning), and academic red-team work (AgentPoison, MINJA, PoisonedRAG, MemoryGraft) demonstrates greater than 80-95% attack success rates against RAG-based agents from small numbers of poisoned documents or query-only interactions, requiring no model retraining. Fourth, "cloud camouflage": attackers route agent-mediated communications through popular, allow-listed SaaS platforms so the AI itself becomes the dispatcher for collaboration posts, ticket updates, or "summaries" that carry sensitive data past egress controls tuned for classic C2 traffic. As a real-world proof of concept, the source article's author built a functioning Chrome credential-stealing capability using AI-driven "vibe coding" platforms (Lovable, Base44, Netlify, Vercel) via an "immersive world" context-manipulation technique -- no manual code was written, illustrating what the author calls a "zero-knowledge threat actor": minimal technical expertise is now needed to assemble offensive capability once an AI agent provides code generation, execution, file-system access, and knowledge-base integration on the attacker's behalf. This tradecraft is corroborated by two independent, contemporaneous disclosures cited in follow-on research: (1) Wiz Research's July 2025 discovery of a critical authentication-bypass flaw in the vibe-coding platform Base44 (subsequently acquired by Wix), in which providing only the non-secret app_id to undocumented registration/OTP-verification endpoints let an attacker fully bypass SSO and app-level auth to access private enterprise apps built on the platform -- patched within 24 hours, no confirmed in-the-wild exploitation; and (2) OX Security's April-May 2026 "Mother of All AI Supply Chains" research, an architectural "RCE-by-design" flaw baked into Anthropic's official MCP SDKs across Python, TypeScript, Java, and Rust, rippling through 150M+ downloads and up to ~200,000 exposed MCP server instances, with verified zero-click prompt-injection-to-RCE chains demonstrated against the Cursor and Windsurf AI coding IDEs and downstream impact confirmed in LangChain, LiteLLM, and IBM LangFlow. Together these findings show the MCP/agent ecosystem's trust model -- connect-once tool vetting, implicit trust of tool responses in the LLM context, and broad default tool permissions -- as a structural weakness attackers across skill levels are actively probing and, in several documented cases, exploiting. Defensive guidance converges on: minimizing and scoping tool permissions to least privilege, hardening and versioning system prompts, validating tool inputs/outputs against strict schemas, enforcing guardrails outside the model (DLP, rate limiting, egress controls), centralizing agent action logging with anomaly detection for unusual tool chaining, and adversarial red-teaming of agent deployments as a recurring practice rather than a one-time gate.

MITRE ATT&CK techniques by tactic:

Defense Evasion

T1027 Obfuscated Files or Information; T1070 Indicator Removal; T1140 Deobfuscate/Decode Files or Information

Exfiltration

T1041 Exfiltration Over C2 Channel; T1567 Exfiltration Over Web Service

Privilege Escalation

T1068 Exploitation for Privilege Escalation

Command and Control

T1071.001 Web Protocols; T1102 Web Service

Discovery

T1082 System Information Discovery; T1083 File and Directory Discovery

Collection

T1119 Automated Collection; T1213 Data from Information Repositories

Initial Access

T1189 Drive-by Compromise; T1190 Exploit Public-Facing Application; T1199 Trusted Relationship; T1566.002 Spearphishing Link

initial-access

T1195.002 Compromise Software Supply Chain

Execution

T1203 Exploitation for Client Execution; T1204.001 Malicious Link

lateral-movement

T1210 Exploitation of Remote Services

Credential Access

T1552 Unsecured Credentials; T1555.003 Credentials from Web Browsers

Impact

T1565.001 Stored Data Manipulation

Resource Development

T1583.006 Web Services; T1585.003 Cloud Accounts; T1588.002 Tool

Reconnaissance

T1592 Gather Victim Host Information

Remediation

Patches

  • Apply MCP SDK vendor patches addressing the OX Security "Mother of All AI Supply Chains" RCE-by-design disclosure across Python/TypeScript/Java/Rust implementations
  • Apply the Base44/Wix fix for the app_id-based registration/OTP authentication-bypass flaw (patched by vendor within 24 hours of disclosure, July 2025)

Immediate actions

  • Inventory every MCP server and third-party AI agent connector in use; remove or disable unreviewed/unsanctioned MCP servers and vibe-coding-platform integrations
  • Restrict agent and MCP tool scopes to least privilege -- disable unrestricted filesystem mounts, full-tenant search, and unbounded outbound web fetch on agent tool chains
  • Require explicit human confirmation for high-risk agent actions (data export, credential access, code execution, external posting)
  • Patch/upgrade MCP SDKs (Python, TypeScript, Java, Rust) to versions addressing the OX Security-disclosed RCE-by-design flaw; audit self-hosted MCP servers for the same pattern
  • Audit any internally built or vendor apps on Base44/Wix for stale unauthenticated registration or OTP-verification endpoints

Workarounds

  • Where MCP servers cannot be immediately patched, isolate them in network-segmented, non-privileged execution contexts and disable tool auto-approval
  • Treat all agent-ingested external content (documents, webpages, tool responses) as untrusted input and pass it through structured-schema validation before it reaches agent tool-invocation logic

Longer-term hardening

  • Deploy DLP, rate limiting, and egress controls scoped specifically to agent/LLM traffic, not just classic user or service-account traffic
  • Centralize agent action and tool-call logging (who/what/when/which tool/what data) with behavioral-anomaly detection for unusual tool-chaining sequences
  • Adopt provenance-aware, trust-scored retrieval for RAG/vector-store content per OWASP Agentic AI Top 10 ASI06 guidance; sanitize and periodically re-validate long-term agent memory
  • Establish an allowlist/registry of approved MCP servers and vibe-coding platforms with a formal security review gate before adoption
  • Institute recurring adversarial red-teaming of production agent deployments (prompt injection, tool-poisoning, memory-poisoning scenarios) as a standing program, not a one-time assessment

Affected and fixed versions

  • Anthropic — Model Context Protocol SDKs (Python, TypeScript, Java, Rust)
    Vulnerable versions: pre-patch releases across all four language SDKs
    Fixed in: patched releases post OX Security disclosure, April-May 2026
  • Wix / Base44 — Base44 vibe-coding platform
    Vulnerable versions: platform state prior to 2025-07-13 fix
    Fixed in: patched 2025-07-13
  • Anysphere — Cursor (AI coding IDE, MCP client)
    Vulnerable versions: versions with unpatched MCP zero-click prompt-injection-to-RCE chain
    Fixed in: unspecified in source reporting
  • Codeium — Windsurf (AI coding IDE, MCP client)
    Vulnerable versions: versions with unpatched MCP zero-click prompt-injection-to-RCE chain
    Fixed in: unspecified in source reporting
  • Multiple — LangChain, LiteLLM, IBM LangFlow (MCP-integrating frameworks)
    Vulnerable versions: versions consuming affected MCP SDKs pre-patch
    Fixed in: unspecified in source reporting

Timeline

  • Wiz Research discovers a critical authentication-bypass flaw in the Base44 vibe-coding platform, exploitable via app_id-only access to undocumented registration/OTP endpoints.
  • Wiz confirms Wix/Base44 has deployed an effective fix for the authentication-bypass flaw.
  • Wix/Base44 officially confirms resolution of the authentication-bypass vulnerability.
  • Wiz Research and multiple outlets (The Hacker News, SecurityWeek, Dark Reading) publish public disclosure of the Base44 authentication-bypass flaw.
  • Etay Maor (Cato Networks) publishes "Living off the AI: The Next Evolution of Attacker Tradecraft" on SecurityWeek, naming prompt-born tool abuse, MCP tooljacking, memory/retrieval poisoning, and vibe-coding-built malware as a converging tradecraft pattern, including a self-built Chrome credential-stealer PoC.
  • OX Security publishes "The Mother of All AI Supply Chains," disclosing an architectural RCE-by-design flaw across Anthropic's MCP SDKs (Python, TypeScript, Java, Rust) affecting up to ~200,000 exposed MCP server instances, with verified zero-click prompt-injection-to-RCE chains in Cursor and Windsurf.
  • SecurityWeek reports on the OX Security MCP supply-chain disclosure, noting downstream impact on LangChain, LiteLLM, and IBM LangFlow.
  • Microsoft Security publishes "Securing AI agents: When AI tools move from reading to acting," providing enterprise defensive guidance for agent-mediated tool abuse consistent with the living-off-the-AI pattern.

Sources cited

Previous and next in threat intel

Drafted by the Threadlinqs research pipeline from 10 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats