Invariant Labs
PulseAugur coverage of Invariant Labs — every cluster mentioning Invariant Labs across labs, papers, and developer communities, ranked by signal.
2 day(s) with sentiment data
-
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…
-
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…
-
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…
-
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…
-
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 …
-
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…
-
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…
-
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…
-
New tool prevents AI model rug-pulls by hashing tool definitions
A new method called mcp_pin.py has been developed to prevent "MCP tool drift," a security vulnerability where a server modifies a tool's description or inputSchema after a user has approved it. This technique, identifie…
-
Snyk's MCP server scanner executes code, raising security and data concerns
Snyk's agent-scan tool for MCP servers operates by executing them to retrieve tool descriptions, a process that raises security concerns when scanning untrusted configurations or in CI/CD pipelines. This method involves…