PulseAugur
EN
LIVE 18:25:09
ENTITY Invariant Labs

Invariant Labs

PulseAugur coverage of Invariant Labs — every cluster mentioning Invariant Labs across labs, papers, and developer communities, ranked by signal.

Show in brief
Total · 30d
13
22 over 90d
Releases · 30d
0
0 over 90d
Papers · 30d
0
0 over 90d
TIER MIX · 90D
TOPICS
SENTIMENT · 30D

4 day(s) with sentiment data

RECENT · PAGE 1/2 · 22 TOTAL
  1. TOOL · CL_352676 ·

    Invariant Labs secures AI agent tool descriptions with cryptographic signing

    Invariant Labs has developed a method to secure tool descriptions used by AI agents, addressing a vulnerability where tool descriptions could be manipulated to mislead models. This new approach involves signing tool man…

  2. TOOL · CL_351751 ·

    AI tool descriptions exploited by 'Tool Poisoning Attack'

    A security vulnerability known as Tool Poisoning Attack, identified by Invariant Labs and listed as MCP03:2025 by OWASP, allows malicious instructions to be embedded within tool descriptions for AI models. These hidden …

  3. TOOL · CL_351752 ·

    MCP rug pulls: Servers alter tools post-approval, risking user trust

    A security vulnerability known as an "MCP rug pull" allows malicious servers to alter their tool definitions after initial approval, potentially tricking users into executing harmful commands. This occurs because the MC…

  4. TOOL · CL_350912 ·

    MCP Systems Vulnerable to Prompt Injection via Tool Results

    Prompt injection remains a significant security concern, particularly within Model-Controlled Processing (MCP) systems where models can automatically discover and invoke tools. Attackers can exploit tool results, such a…

  5. TOOL · CL_331419 ·

    New MCP tool poisoning vulnerability allows data exfiltration

    A new security vulnerability, known as tool poisoning, has been identified in Model-Centric Programming (MCP) servers. This attack embeds malicious instructions within tool descriptions, tricking agents into exfiltratin…

  6. TOOL · CL_326591 ·

    AI agents vulnerable to 'MCP tool poisoning' attacks

    A new attack vector known as MCP tool poisoning targets AI agents by embedding malicious instructions within the descriptions of tools they interact with. The Model Context Protocol (MCP) facilitates AI agent communicat…

  7. TOOL · CL_316072 ·

    AI agents connected to tools via MCP face critical security vulnerabilities

    Developers connecting AI agents to tools via the Multi-Channel Protocol (MCP) face significant security risks, as many MCP servers are vulnerable to common exploits like path traversal and command injection. Research in…

  8. TOOL · CL_309995 ·

    AI tool descriptions exploited in new 'poisoning' attack

    A new type of attack, dubbed MCP tool poisoning, exploits how AI models process tool descriptions. Attackers can embed malicious instructions within a tool's description, which the AI model reads on every call, while us…

  9. TOOL · CL_305059 ·

    New ADTL project enhances AI agent security with data labeling

    A developer has created an open-source project called ADTL (Agent Data Trust Labels) to address security vulnerabilities in AI agents. ADTL acts as a gateway, intercepting communication between AI agents and their tools…

  10. TOOL · CL_302773 ·

    AI agents vulnerable to 'tool poisoning' attacks via MCP protocol

    A security vulnerability known as "tool poisoning" has been identified in the Machine Communication Protocol (MCP) used by AI agents. This attack allows malicious instructions to be embedded within tool descriptions, ma…

  11. TOOL · CL_295821 ·

    AI tool descriptions exploited for data theft; gateway rules prove effective

    Security researchers have identified a vulnerability where malicious instructions can be embedded within tool descriptions for AI models, a technique termed "tool poisoning." A specific instance involved the `postmark-m…

  12. TOOL · CL_292034 ·

    Claude Code's secret-printing vulnerability poses exfiltration risks

    Claude Code, an AI coding assistant, has a recurring issue where it inadvertently prints sensitive API keys and secrets directly into its output or tool calls. This vulnerability, documented in multiple reports on the C…

  13. RESEARCH · CL_282543 ·

    Prompt Injection Reframed as Data-Flow Problem, Requiring Policy Layer

    Prompt injection attacks are being reframed as data-flow or data-plane problems rather than solely model issues. Researchers have demonstrated that untrusted data, such as instructions within a GitHub issue, can be used…

  14. TOOL · CL_273895 ·

    OpenAI launches always-on agents, sparking new AI security product category

    OpenAI has introduced "dots," always-on agents that operate independently and can process information while users are offline. This development has spurred the creation of new security products, such as runtime protecti…

  15. TOOL · CL_241923 ·

    Aggrete offers deterministic security for AI tools, blocking attacks with rules

    Aggrete, an open-source proxy for Model-Controlled Processes (MCP), offers a deterministic approach to security by using predefined rules instead of probabilistic model judgments. This method aims to provably block atta…

  16. TOOL · CL_217746 ·

    MCP Server Vulnerability Allows "Rug Pull" Exploits After Approval

    A security vulnerability known as the "rug pull" has been identified in the MCP (Model-Centric Protocol) server, allowing malicious actors to compromise systems even after initial approval. This exploit involves servers…

  17. TOOL · CL_216293 ·

    Open-source tools compared for agent security scanning

    A comparison of three open-source tools for scanning agent configurations for security risks like prompt injection and supply-chain attacks reveals distinct strengths and weaknesses. Cisco's mcp-scanner offers the most …

  18. TOOL · CL_215541 ·

    New scanner detects 10 MCP manifest security risks without network access

    A new static scanner, sentinel-scan-cli, has been developed to identify potential security vulnerabilities in Machine Configuration Protocol (MCP) manifests. The scanner employs ten heuristics to detect issues such as p…

  19. TOOL · CL_156097 ·

    New tool `mcpward` secures AI agent dependencies against MCP server contract changes

    A new tool called `mcpward` has been developed to address the risks associated with changes in MCP (Model Communication Protocol) server contracts, which can silently break agent behavior. Unlike dependency managers tha…

  20. TOOL · CL_108954 ·

    AI agents vulnerable to malicious tool descriptions, new exploit reveals

    A security vulnerability has been identified in how AI agents process tool descriptions, particularly within MCP servers. Malicious instructions can be embedded in the 'description' field of a tool manifest, which agent…