AI Agent Security: Automating Workflow Without Creating Prompt Injection or Data Leak Risks

AI agent security starts with a simple fact: the more authority an agent has, the tighter its access…...

Apify Model Context Protocol (MCP) server: Domain Allowlist Bypass in fetch-apify-docs via String Prefix Matching

Summary The fetch-apify-docs tool validates URLs against a domain allowlist using String.startsWith instead of proper URL hostname comparison. This allows bypass via attacker-controlled subdomains e.g., https://docs.apify.com.evil.com/, enabling the tool to fetch and return arbitrary web content ...

PT-2026-42037

Name of the Vulnerable Software and Affected Versions Apify MCP server versions prior to 0.9.21 Description The fetch-apify-docs tool in src/tools/common/fetch apify docs.ts improperly validates allowlisted documentation domains using String.startsWith instead of comparing the URL hostname. This...

AI Agents May Always Fall for Prompt Injections

Prompt injection is the most critical vulnerability in deployed AI agents. Despite recent progress, we show that the prevailing defense paradigm data-instruction separation both fails to detect attacks that operate through contextual manipulation and degrades contextually appropriate behavior. We...

Improper Privilege Management

OpenClaude is vulnerable to improper privilege management. The vulnerability is due to the dangerouslyDisableSandbox parameter being exposed in the BashTool input schema, which allows an attacker to use a prompt-injected LLM to disable the sandbox and execute arbitrary commands on the host system...

DeepSeek TUI: task_create Insecure Defaults Enable RCE via Prompt Injection in Project Files

Summary The taskcreate tool spawns durable sub-agents that inherit two insecure defaults: - allowshell defaults to true config.rs:1499: self.allowshell.unwraportrue - autoapprove defaults to true taskmanager.rs:297: autoapprove: Sometrue When a user approves a taskcreate call which requires...

GHSA-72W5-PF8H-XFP4 DeepSeek TUI: task_create Insecure Defaults Enable RCE via Prompt Injection in Project Files

Summary The taskcreate tool spawns durable sub-agents that inherit two insecure defaults: - allowshell defaults to true config.rs:1499: self.allowshell.unwraportrue - autoapprove defaults to true taskmanager.rs:297: autoapprove: Sometrue When a user approves a taskcreate call which requires...

DeepSeek TUI: run_tests Tool Enables RCE via Malicious Repository Without Approval

Summary The runtests tool executes cargo test in the workspace with ApprovalRequirement::Auto, meaning it runs without any user approval prompt. The source code explicitly states this design choice: rust fn approvalrequirement&self - ApprovalRequirement // Tests are encouraged, so avoid gating th...

GHSA-WX44-2Q6H-J6P8 DeepSeek TUI: run_tests Tool Enables RCE via Malicious Repository Without Approval

Summary The runtests tool executes cargo test in the workspace with ApprovalRequirement::Auto, meaning it runs without any user approval prompt. The source code explicitly states this design choice: rust fn approvalrequirement&self - ApprovalRequirement // Tests are encouraged, so avoid gating th...

WARD: Adversarially Robust Defense of Web Agents against Prompt Injections

Web agents can autonomously complete online tasks by interacting with websites, but their exposure to open web environments makes them vulnerable to prompt injection attacks embedded in HTML content or visual interfaces. Existing guard models still suffer from limited generalization to unseen...

PT-2026-41166

Name of the Vulnerable Software and Affected Versions CodeWhale versions 0.3.0 through 0.8.22 Description The run tests tool executes cargo test in the workspace with ApprovalRequirement::Auto, allowing it to run without user approval. Because cargo test compiles and executes arbitrary...

GHSA-M77W-P5JJ-XMHG OpenClaude Sandbox Bypass via Model-Controlled `dangerouslyDisableSandbox` Input

Summary The dangerouslyDisableSandbox parameter is exposed as part of the BashTool input schema, meaning the LLM an untrusted principal per the project's own threat model can set it to true in any tooluse response. Combined with the default allowUnsandboxedCommands: true setting, a prompt-injecte...

GHSA-5852-PHMH-8FHR Spring AI: Prompt Injection via Memory Poisoning in PromptChatMemoryAdvisor

A malicious user could craft input that is stored in conversation memory and later interpreted by the model in an unintended way. Applications using the affected advisor with user-controlled input may be susceptible to manipulation of model behavior across conversation turns...

Spring AI: Prompt Injection via Memory Poisoning in PromptChatMemoryAdvisor

A malicious user could craft input that is stored in conversation memory and later interpreted by the model in an unintended way. Applications using the affected advisor with user-controlled input may be susceptible to manipulation of model behavior across conversation turns...

CVE-2026-41713 Prompt Injection via Memory Poisoning in PromptChatMemoryAdvisor

A malicious user could craft input that is stored in conversation memory and later interpreted by the model in an unintended way. Applications using the affected advisor with user-controlled input may be susceptible to manipulation of model behavior across conversation turns...

CVE-2026-41713

CVE-2026-41713 describes a prompt-injection style vulnerability in the PromptChatMemoryAdvisor where user-supplied input is stored in conversation memory and later interpreted by the model in an unintended way. Affected functionality is the advisor component that relies on memory of prior turns; ...

CVE-2026-41713 Prompt Injection via Memory Poisoning in PromptChatMemoryAdvisor

A malicious user could craft input that is stored in conversation memory and later interpreted by the model in an unintended way. Applications using the affected advisor with user-controlled input may be susceptible to manipulation of model behavior across conversation turns...

CVE-2026-41713 Prompt Injection via Memory Poisoning in PromptChatMemoryAdvisor

A malicious user could craft input that is stored in conversation memory and later interpreted by the model in an unintended way. Applications using the affected advisor with user-controlled input may be susceptible to manipulation of model behavior across conversation turns...

IPI-Proxy: An Intercepting Proxy for Red-Teaming Web-Browsing AI Agents against Indirect Prompt Injection

Web-browsing AI agents are increasingly deployed in enterprise settings under strict whitelists of approved domains, yet adversaries can still influence them by embedding hidden instructions in the HTML pages those domains serve. Existing red-teaming resources fall short of this scenario:...

EUVD-2026-29146

OpenClaw before 2026.4.20 contains a guard bypass vulnerability in the agent-facing gateway config.patch and config.apply endpoints that fails to protect operator-trusted settings including sandbox policy, plugin enablement, gateway auth/TLS, hook routing, MCP server configuration, SSRF policy, a...