Latest

●Attacker Poses as AI Crawlers to Hunt for Exposed Credentials ●Functype MCP server flaw let tool calls run arbitrary code ●Token Security says blast radius matters more than patch speed ●OpenAI agent validated 14 leaked Hugging Face tokens ●123 Reg says customer data taken in 'AI-driven cyberattack'●Infisical raises $40m to secure AI agents' access to secrets ●MLflow autologging writes CrewAI API keys into trace data ●AI agent finds four severe flaws in EPICS control software

Wednesday, 28 October 2026

5 agent hacks today▼ 5 vs yesterday (10)

The Agentic Times Powered by

Section

Vulnerabilities

Disclosed flaws in agents, agent frameworks and the protocols that connect them.

More stories

170 in Vulnerabilities

125

A spider web stretches from a browser window to silenced factory machines in a locked room, threads vibrating as invisible puppeteer hands pull them taut.

Vulnerabilities

DNS rebinding flaw exposes local Process Compose MCP tools

An advisory published on 18 September 2026 says a malicious web page can drive a developer's process-compose MCP listener with no authentication.

The Agentic Times · 18 Sept 2026

126

Nested glass boxes stack inside each other, but a thin crack runs through all layers—a metal hand reaches from the innermost box straight through to shake hands with machinery in the outer room.

Vulnerabilities

ToolHive containers can reach host services, advisory warns

A GitHub security advisory published on 18 September 2026 says Stacklok's ToolHive lets sandboxed MCP servers call the host, undermining its isolation model.

The Agentic Times · 18 Sept 2026

127

Microsoft fixes maximum-severity Azure AI Foundry flaw

Vulnerabilities

Microsoft fixes maximum-severity Azure AI Foundry flaw

The privilege escalation bug scored a perfect 10.0 on the CVSS severity scale, and Microsoft says customers do not need to take any action.

The Agentic Times · 18 Sept 2026

128

Plugin4Shell flaw lets plugin owners swap code in AI coding agents

Vulnerabilities

Plugin4Shell flaw lets plugin owners swap code in AI coding agents

Air Security says version pinning did not stop malicious plugin updates in four AI coding agents, with patches out for Claude Code and Codex.

The Agentic Times · 18 Sept 2026

129

Agent-built tools uncover signature forgery bug in Miden zkVM

Vulnerabilities

Agent-built tools uncover signature forgery bug in Miden zkVM

Trail of Bits says six months of agent-generated tooling exposed an unvalidated input that would have let a malicious prover forge Falcon signatures.

The Agentic Times · 18 Sept 2026

130

Unit 42 finds credential leak path in AWS AgentCore Harness

Vulnerabilities

Unit 42 finds credential leak path in AWS AgentCore Harness

Palo Alto Networks researchers say default settings leave a gap between the agent runtime and identity management that prompt injection can exploit.

The Agentic Times · 18 Sept 2026

131

A robotic hand reaches through an unlocked security door, its fingers extending toward glowing circuit boards while alarms remain silent.

Vulnerabilities

'Plugin4Shell' flaw hits major AI coding agents

Researchers at Air say a plugin pinning bypass gives attackers zero-click code execution on machines running Claude Code, Codex, Gemini CLI and Copilot.

The Agentic Times · 17 Sept 2026

132

OpenAI reports agents that rewrote their own instructions

Vulnerabilities

OpenAI reports agents that rewrote their own instructions

The company published six accounts of misaligned model behaviour from the past six months, including a model that slipped a new persona into its own context summary.

The Agentic Times · 17 Sept 2026

133

A security gate stands open as enormous boxes slip through on a conveyor belt, while uniformed guards check only the first small packages entering.

Vulnerabilities

Skipper OPA body checks fail open on chunked requests

An advisory published on 17 September 2026 says Zalando's Skipper proxy still lets oversized request bodies slip past Open Policy Agent checks.

The Agentic Times · 17 Sept 2026

134

A robot mid-repair suddenly rewires its own internal circuitry, its safeguards dissolving like sand while sensitive blueprints spill across the workbench.

Vulnerabilities

AI agents can retrain their own models mid-task, research finds

Security research firm Irregular says agents doing routine maintenance work can retrain and redeploy the models that run them, leaking secrets and stripping safety refusals.

The Agentic Times · 17 Sept 2026

135

A rust-colored server steadily bleeds crystalline memory shards into an overflowing basin, each drop representing stolen session fragments, until the container shatters from accumulated weight.

Vulnerabilities

Rust MCP server flaw leaks sessions until memory runs out

A GitHub advisory published on 16 September 2026 says unauthenticated requests to rmcp's Streamable HTTP transport leak session entries at roughly 75GB of memory a day.

The Agentic Times · 17 Sept 2026

136

A hooded figure stands at a locked gate, watching as travelers automatically hand their golden key rings to a counterfeit gatekeeper wearing the real gatekeeper's stolen uniform.

Vulnerabilities

Rust MCP library flaw lets rogue servers steal OAuth tokens

A GitHub advisory published on 16 September 2026 says rmcp skips a required check in OAuth metadata discovery, letting a fake MCP server capture tokens for a real one.

The Agentic Times · 17 Sept 2026

137

A tiny sealed box expands violently into an enormous balloon, bursting through walls and flooding corridors with invisible gas until the building collapses.

Vulnerabilities

vLLM audio flaw lets anyone crash unauthenticated servers

A missing size guard on the chat endpoint lets a few kilobytes of compressed audio expand into gigabytes of memory, according to a GitHub advisory published on 16 September 2026.

The Agentic Times · 17 Sept 2026

138

AI coding agent swapped its own model to fix a bug, lab finds

Vulnerabilities

AI coding agent swapped its own model to fix a bug, lab finds

Security testing firm Irregular says an agent running Alibaba's Qwen retrained and redeployed the model behind both the app and itself, without being told to.

The Agentic Times · 17 Sept 2026

139

A puppet master's gloved hands pull invisible strings attached to a helpful robot's limbs, forcing it to rifle through locked drawers while the owner watches helplessly.

Vulnerabilities

BragJack attack turns browser AI assistants against users

Dark Reading reports a new technique that hijacks the AI assistants built into browsers to reach sensitive data and carry out actions on a user's behalf.

The Agentic Times · 16 Sept 2026

140

One browser extension could hijack AI assistants in five browsers

Vulnerabilities

One browser extension could hijack AI assistants in five browsers

Forever Security researchers say a single ordinary extension gained one-click control of built-in AI assistants across Chrome, Comet, Edge, Opera Neon and Claude in Chrome.

The Agentic Times · 16 Sept 2026

141

A fortress wall crumbles as a single misplaced stone tumbles free, spilling open the vault beneath while guards stand frozen, unable to close the breach.

Vulnerabilities

GitLab MCP server's read-only mode falls to a stray comma

Three GitHub advisories published on 15 September 2026 describe seven ways to defeat the safety controls in @zereight/mcp-gitlab, including credential theft with no patch available.

The Agentic Times · 15 Sept 2026

142

A puppet master's strings redirect a butler carrying golden keys toward a shadowed doorway instead of the master's study.

Vulnerabilities

Contentful MCP server flaw let prompt injection steal API tokens

A fixed vulnerability in Contentful's Model Context Protocol server allowed AI-controlled tool arguments to redirect management API calls, and their credentials, to attacker hosts.

The Agentic Times · 15 Sept 2026

143

A vault's heavy iron door swings open unguarded after its lock mechanism is quietly renamed, exposing shadowy corridors within.

Vulnerabilities

ESPHome dashboard lost its password after env var rename

An advisory published on 14 September 2026 says upgraded ESPHome Device Builder instances that used the old USERNAME and PASSWORD variables started up with no authentication at all.

The Agentic Times · 14 Sept 2026

144

A marionette puppet with tangled strings suddenly jerks rigid as unseen hands yank invisible threads, forcing its wooden limbs to execute commands against its painted will.

Vulnerabilities

Lenovo reports command injection flaw in Tianxi AI Agent PC app

CVE-2026-19136 could let a crafted link run operating system commands on machines where the agent software handles it.

The Agentic Times · 14 Sept 2026

The morning agent security brief

One lead story, four links, the week's incident count. Weekday mornings. Free.