Ajar: Measuring Open Privilege in Agent Defenses

A language model agent acts through the tools it is given. The data it reads while working on a task can redirect what it does with those tools. A growing set of techniques for safe and secure agent execution therefore sits between the agent and its tools, aiming to enforce access control,...

CVE-2026-77521

MaxKB is an open-source enterprise AI assistant. Prior to 2.10.5-lts , assistants with a tool, MCP tool, skill, or sub-application use SandboxShellBackend , which exposes an execute shell tool without excluding it and omits execute from interrupt_on, meaning no human approval is required before e...

CVE-2026-77521 MaxKB: Prompt-injectable agent can lead to command execution

MaxKB is an open-source AI assistant for enterprise. Prior to version 2.10.5-lts, assistants with a tool, MCP tool, skill, or sub-application use SandboxShellBackend, which exposes an execute shell tool without excluding it and omits execute from interrupton, so human approval is not required...

CVE-2026-77521 MaxKB: Prompt-injectable agent can lead to command execution

MaxKB is an open-source AI assistant for enterprise. Prior to version 2.10.5-lts, assistants with a tool, MCP tool, skill, or sub-application use SandboxShellBackend, which exposes an execute shell tool without excluding it and omits execute from interrupton, so human approval is not required...

CVE-2026-77521 MaxKB: Prompt-injectable agent can lead to command execution

MaxKB is an open-source AI assistant for enterprise. Prior to version 2.10.5-lts, assistants with a tool, MCP tool, skill, or sub-application use SandboxShellBackend, which exposes an execute shell tool without excluding it and omits execute from interrupton, so human approval is not required...

ironcurtain memory-mcp-server/v0.2.1

IronCurtain A secure runtime for autonomous AI agents, where security policy is derived from a human-readable constitution. When someone writes "secure," you should immediately be skeptical.What do we mean by secure? !WARNING Research Prototype. IronCurtain is an early-stage research project...

Decoding Guardrails: XAI-Guided Perturbation Analysis of Prompt Injection Detection

Large language models LLMs are increasingly deployed in production systems, raising concerns about their exposure to adversarial manipulation through prompt injection and jailbreak attacks. Classifier-based guardrails, such as Prompt Guard 2, are widely used as a first line of defense against suc...

PT-2026-96257

Name of the Vulnerable Software and Affected Versions MaxKB versions prior to 2.10.5-lts Description Assistants utilizing a tool, MCP tool, skill, or sub-application use the SandboxShellBackend, which exposes an execute shell tool. Because the execute function is omitted from the interrupt on...

firefox-devtools-mcp v0.10.3

Firefox DevTools MCP Model Context Protocol server for automating Firefox via WebDriver BiDi through Selenium WebDriver. Works with Claude Code, Claude Desktop, Cursor, Cline and other MCP clients. Repository: https://github.com/mozilla/firefox-devtools-mcp Note : This MCP server requires a local...

agent-scan v0.6.4-snapshot-6e2d290-1638

Snyk Agent Scan Discover and scan agent components on your machine for prompt injections and vulnerabilities including agents, MCP servers, skills. Note: We don't publish an npm package for Agent Scan. Install it via uvx or as a standalone binary. Note: CLI output is experimental and subject to...

CVE-2026-93841 vLLM through 0.29.0 Adjacent-Request Sampler State Corruption via Unvalidated Prompt Token IDs

vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton bincountkernel where prompt token IDs index the penalty prompt-presence bitset without bounds checking against vocabulary size. Attackers can submit multimodal audio requests with tokens equal to vocabulary size, causing...

CVE-2026-93839 LightLLM through 1.2.0 Missing Authentication in PD Master /pd_register WebSocket Endpoint

LightLLM through 1.2.0 contains an authentication bypass vulnerability in the /pdregister WebSocket endpoint that allows unauthenticated attackers to register arbitrary nodes by supplying crafted JSON without peer address validation. Attackers can disclose full user prompts routed to their socket...

CVE-2026-93839

LightLLM (by ModelTC) through version 1.2.0 contains a missing authentication vulnerability in the /pd_register WebSocket endpoint of the PD Master component. An unauthenticated attacker can register arbitrary nodes by supplying crafted JSON that bypasses peer address validation. Successful explo...

PYSEC-2026-4014 Command injection via --prompt in activate.bat (batch activator)

BatchActivator.quote returned its input unchanged, the only activator with no escaping at all. --prompt, the VIRTUALENVPROMPT environment variable, and the config file all set the prompt, and activate.bat writes it straight into @set "VAR=value". A prompt containing a double quote closes that...

Command injection via --prompt in activate.bat (batch activator)

BatchActivator.quote returned its input unchanged, the only activator withno escaping at all. --prompt, the VIRTUALENVPROMPT environment variable,and the config file all set the prompt, and activate.bat writes it straightinto @set "VAR=value". A prompt containing a double quote closes that...

PYSEC-2026-4012 virtualenv writes prompt values into pyvenv.cfg without sanitizing line boundaries, allowing configuration injection

pyvenv.cfg is a line-based format with no escape syntax. PyEnvCfg.write wrote values verbatim, while PyEnvCfg.readvalues parses the file with str.splitlines. A value containing a line boundary therefore became additional configuration lines, and because reading is last-wins, the injected keys...

virtualenv writes prompt values into pyvenv.cfg without sanitizing line boundaries, allowing configuration injection

pyvenv.cfg is a line-based format with no escape syntax. PyEnvCfg.writewrote values verbatim, while PyEnvCfg.readvalues parses the file withstr.splitlines. A value containing a line boundary therefore becameadditional configuration lines, and because reading is last-wins, the injectedkeys replace...

jfp-kernel-sandbox-poc

JFP Kernel Sandbox bwrap PoC A Proof of Concept PoC demon...

CLEANSTART-2026-JK52932 expansion of '\h' in the prompt string in bash 4

Security vulnerability affects the bash package. The expansion of '\h' in the prompt string in bash 4...

PT-2026-105773

pyvenv.cfg is a line-based format with no escape syntax. PyEnvCfg.write wrote values verbatim, while PyEnvCfg. read values parses the file with str.splitlines. A value containing a line boundary therefore became additional configuration lines, and because reading is last-wins, the injected keys...