Threat Library

PraisonAI's local MCP HTTP Stream server uses a flawed startswith() check to validate the HTTP Origin header, allowing an attacker-controlled domain like 'localhost.evil.example' to pass as if it were 'localhost'. A malicious webpage visited by a victim can exploit this to send unauthenticated tools/call requests directly to the victim's local MCP server, invoking any exposed tool without an API key.

QWED, an infrastructure component used to verify LLM/agent tool outputs and math expressions before production execution, passes untrusted expressions directly to SymPy's parse_expr() without restricting namespaces, allowing arbitrary Python code execution in the API server. Because signup and API key issuance are open by default, any external attacker can register a tenant, obtain a valid API key, and trigger remote code execution via the math verification endpoints, fully compromising the server and other tenants in shared deployments.

browse-mcp, an MCP server that gives agents headless-browser capabilities, fails to validate caller-controlled save paths in its browser_download, browser_save_state, and browser_load_state tools, allowing arbitrary file writes anywhere the process can reach. An attacker who controls a webpage the agent visits can use indirect prompt injection to steer the agent into writing attacker-controlled content to sensitive locations like ~/.bashrc or cron files, potentially achieving full host code execution. A separate flaw in the force_fetch fallback also bypasses the configured origin allowlist entirely.

This article is a general commentary on the security vendor landscape at Black Hat 2026, noting that AI and agents are heavily featured in vendor marketing. It contains no technical details about a vulnerability, exploit, or attack against AI agents, frameworks, or protocols. This is industry analysis, not a security threat report.

This item is a press/blog announcement from Unit 42 summarizing a broader research report on AI-enabled malware trends, not a specific technical disclosure of an agent-related vulnerability or attack. There is no actionable detail about a concrete AI-agent, MCP, or A2A threat in the provided data, so it does not constitute a genuine standalone security issue for this analysis.

This item is an academic research paper describing a new defensive fine-tuning technique (SecOPD) that reduces LLM susceptibility to adaptive prompt injection attacks, rather than describing a new vulnerability, active exploit, or attack technique. It is a mitigation contribution, not a threat disclosure, so it is classified as low severity from a threat-intelligence standpoint, though it is highly relevant context for understanding the ongoing prompt injection problem in agentic systems.

This is an academic research paper describing a new defensive technique (AEGIS) against indirect prompt injection attacks in LLM-based agents, not a description of an active exploit or newly discovered vulnerability. It proposes using instruction-sensitive projectors and multi-layer consensus signals to detect malicious injected instructions while reducing over-refusal of legitimate inputs. The underlying threat class (indirect prompt injection) is well-known and already covered by existing threat intelligence; this submission itself introduces no new attack technique.

This is not a vulnerability disclosure but a research paper describing AgentFlow, a defensive framework that enforces data-flow policies across LLM agent tool calls and delegation boundaries to prevent prompt-injection-driven data exfiltration and unsafe cross-agent authority propagation. The work reports strong reductions in confirmed compromise rates across several established agent security benchmarks (AgentDojo, AgentDyn, ASB, InjecAgent, BIPIA, AgentHarm, MCPTox). Severity is rated low because the raw data describes a proposed mitigation, not an active threat, exploit, or newly discovered weakness.

The Continue CLI's headless/auto mode relies solely on an incomplete denylist to block dangerous shell commands, missing paths like /home, /root, /var, and failing to properly handle variable expansion such as $HOME. An attacker who can inject instructions into content the agent reads (web pages, repo files, issues) can trigger destructive commands like recursive deletion, causing unattended agent runs to destroy user data.

This is a research paper describing a defensive framework (AEGIS), not an active exploit or newly disclosed vulnerability. It highlights a known class of risk where MCP agents can request excessive resources (e.g., huge search radii, long videos) across multiple modalities, potentially degrading service, and proposes a policy-enforcement layer to mitigate it.

This is a defensive research paper describing ClawSentry, a security gateway that monitors LLM agents to detect and block malicious third-party skills/tools before and during execution. It is not an active exploit or vulnerability disclosure, but rather a proposed mitigation framework benchmarked against known attack scenarios (SkillInject, SkillsSafety). Severity is low from a threat perspective since this describes protective tooling, though it implicitly confirms that unprotected agent runtimes have high attack success rates (33.5-49.7%) against malicious skill injection.

This is a defensive research paper, not an active threat or exploit disclosure. It proposes TraceGrant, a governance framework that mitigates indirect prompt injection in networked LLM agents by binding tool execution to a pre-established 'Contract' derived from trusted user intent and verifying task completion against actual results. The underlying risk it addresses (indirect prompt injection redirecting tool calls) is well-known and already covered by existing threat models; no new vulnerability or attack technique is introduced here.

This is an academic research paper (not an active exploit) demonstrating that persistent agent memory systems are highly vulnerable to simple, plainly worded false statements: poisoning just 1.2% of a memory corpus dropped answer accuracy from 0.85 to 0.30. The authors show that both write-time content screening and provenance-weighted retrieval defenses fail to catch these low-effort attacks, exposing a fundamental architectural gap rather than a specific software bug.

This item is a blog post by Simon Willison discussing best practices for reviewing and verifying changes made by AI coding agents, not a security vulnerability or threat report. There is no indication of prompt injection, tool poisoning, or any exploitable weakness in agents or protocols.

The CKAN MCP Server's earlier SSRF fix only blocked the literal hostname 'localhost' and dotted IPv4/bracketed IPv6 loopback literals, but missed alternate hostname aliases like 'ip6-localhost' and 'ip6-loopback' that resolve to loopback addresses. A remote caller invoking CKAN tools (e.g. ckan_package_search, sparql_query) with a crafted server_url/base_url parameter could bypass the filter and force the server to make requests to internal or loopback network locations. This is a moderate-severity SSRF bypass affecting a specific tool integration, not a novel or high-impact agentic exploitation technique.

Infracost, a cost-analysis tool used alongside AI coding agents and CI/CD pipelines, could be tricked into sending a configured Terraform Cloud or registry token to an attacker-controlled host. This happens because the tool derives the destination hostname from untrusted Terraform input (e.g., in a pull request) rather than validating it against the trusted host configuration. The issue is fixed in version 0.10.45.

Infracost's template functions used by AI coding agents and CI/CD pipelines fail to properly resolve intermediate directory symlinks, allowing a malicious repository to trick the tool into reading arbitrary files accessible to the CI runner, including secrets. The leaked content can then be surfaced in the Infracost dashboard or pull request comments, creating a data exfiltration path. This is a classic path traversal/symlink vulnerability in a tool commonly invoked by AI agents and CI automation, fixed in version 0.10.45.

Omnigent, an open-source AI agent orchestration framework, fails to validate the os_env.cwd field in uploaded agent bundles, allowing an authenticated user to supply an absolute or traversal path. When the runner has no configured workspace root, this attacker-controlled path becomes the trusted root for file and shell tools, letting an agent read, write, or execute outside its intended sandbox and potentially exfiltrate host secrets.

Omnigent, an open-source AI agent orchestration framework, contains a shell-command parser that fails to recognize certain command syntax (combined flags, timing wrappers, command substitution, background operators). This causes the framework's git/gh write allowlist and workspace confinement policies to silently abstain rather than block, letting an agent push to unauthorized repositories or branches or escape its intended working directory. A prompt-injected or otherwise manipulated coding agent could exploit this gap to perform unauthorized writes outside its granted scope.

Omnigent, an open-source AI agent orchestration framework, fails to validate the dotted Python path an authenticated user supplies for a tool callable when uploading an agent bundle. This allows a low-privilege authenticated user to point a tool at dangerous built-ins like subprocess.check_output, achieving arbitrary command execution with the runner's permissions. The result is full compromise of the runner process, including credentials, environment variables, workspace data, and internal service access.