Latest

A locksmith's salt crystals scatter across a shared hallway floor after a guest uses the master key, leaving fresh footprints visible to the next tenant arriving.

vLLM drops cache salt in tool turns, leaking prompt hints

An advisory published on 6 October 2026 says vLLM silently discards a caller's cache_salt after tool calls, re-opening a cross-tenant prefix-cache oracle.

6 Oct 2026

A puppet master's strings slip through a hastily patched fence, revealing a second hidden gap where commands still whisper through unguarded doors.

Vulnerabilities

simple-git flaw lets crafted clone arguments run shell commands

A GitHub advisory published on 5 October 2026 says the npm package's safety denylist misses Git config includes, and the pending fix misses a variant.

6 Oct 2026

A locksmith hands a stranger a key cut from someone else's lock, unknowingly pairing mismatched tumblers that now open doors to restricted rooms.

Vulnerabilities

PyJWT accepts mismatched Ed25519 key pairs in JWK imports

A flaw disclosed on 5 October 2026 lets a crafted JSON Web Key pair one user's public key with an attacker's private key, breaking DPoP token binding.

6 Oct 2026

A single thread pulled from a spinning loom causes the entire weaving machine to collapse into tangled chaos, destroying hours of previous work.

Vulnerabilities

Six vLLM advisories show single requests can kill the engine

Advisories published on 5 October 2026 describe denial-of-service and cross-request integrity flaws in vLLM versions up to and including 0.25.1.

6 Oct 2026

A mechanical hand reaches through a locked gate into a filing cabinet, extracting sealed envelopes despite the chain barrier intended to block it.

Vulnerabilities

OpenClaw chat plugins let agents read off-limits channels

A GitHub security advisory published on 5 October 2026 says read actions in four OpenClaw messaging plugins could ignore operator allowlists.

6 Oct 2026

A vault's iron door swings open as someone inserts a single master key, revealing rows of glowing credential tokens spilling into shadow.

Vulnerabilities

Langflow flaw lets attackers decrypt stored API keys

A critical weakness in how Langflow derives its encryption key means anyone with the instance secret can read every stored credential, according to a GitHub advisory.

5 Oct 2026

A fortress of interlocking gears suddenly sprouts three gaping holes, allowing shadowy hands to reach through and rewire the machine's inner workings at will.

Vulnerabilities

Three critical flaws let attackers run code on Langflow servers

Advisories published on 5 October 2026 cover a prompt injection path to code execution, an authenticated remote code execution bug and a webhook authentication bypass.

5 Oct 2026

Incidents

South Korea opens inquiry into bank breaches linked to AI

The Financial Services Commission called an emergency meeting after a run of cyberattacks on financial institutions, with investigators weighing whether AI-powered techniques were used.

5 Oct 2026

Vendors

GitGuardian pitches credential sprawl as an AI agent problem

A vendor article published on 5 October 2026 argues that machine and AI identities are widening the credential layer faster than security teams can map it.

5 Oct 2026

Vendors

Apple to tighten macOS Full Disk Access over AI agent risk

The company says some developers are using the permission in ways that expose files, mail, messages and browsing history without users fully knowing.

5 Oct 2026

A robot's mechanical arm spins dials frantically, stacking gold coins higher and higher until a steel wall slams down, halting it mid-motion.

Vendors

Runaway agents revive the case for hard spending caps

A widely read developer blog argues cloud providers should cut off usage by default once a budget is hit, as coding agents make it trivial to spin up costly services.

4 Oct 2026

Vendors

Gemini may get broad access to files and apps on macOS

BleepingComputer reports that Google's assistant could soon open apps, read any file and act on a Mac without asking permission each time.

4 Oct 2026

Incidents

Attackers exploit AI-found flaw in Rejetto file server within a day

A critical authentication bypass in Rejetto HTTP File Server, found with Anthropic's Mythos model, drew real-world attacks the day after disclosure, VulnCheck says.

3 Oct 2026

A puppet master's hands steady a marionette mid-leap over a cliff edge, invisible safety rails glowing beneath its wooden limbs.

Vendors

doxx.net raises $38m for AI agent guardrail platform

SecurityWeek reported on 3 October 2026 that the company's new ADN platform aims to stop agents going wrong while acting under a user's authority.

3 Oct 2026

A vault door stands ajar while invisible hands rifle through filing cabinets; a puppet's strings lead toward shadowy corridors inside fortress walls.

Vulnerabilities

SiYuan patches two AI agent tool flaws in note-taking app

Advisories published on 2 October 2026 say SiYuan's MCP tools could read any local file and be tricked into reaching internal network addresses.

2 Oct 2026

A suit of armor with an open visor accepts weapons from any stranger's hand, then fights with the wielder's own sword.

Vulnerabilities

Headroom AI proxy accepts WebSocket requests from any web page

A GitHub advisory published on 2 October 2026 says the Headroom proxy skips Origin checks and attaches the operator's OpenAI key to attacker-sent requests.

2 Oct 2026

A puppet master's strings bypass a security gate, commanding a marionette to execute invisible commands through invisible doorways.

Vulnerabilities

A2UI flaw let agent-supplied buttons run JavaScript in the browser

An advisory published on 2 October 2026 says @a2ui/web_core passed agent-controlled URLs straight to window.open, with no check on the URI scheme.

2 Oct 2026

A factory worker pulls debug papers from a malfunctioning machine, revealing handwritten passwords and keys scattered across the floor in plain sight.

Vulnerabilities

Trigger.dev CLI debug logs print environment secrets in plain text

A GitHub security advisory published on 2 October 2026 says the workflow platform's command-line tool leaked resolved secrets during dry-run deployments.

2 Oct 2026

A gleaming control room's central console sits unguarded as shadowy hands reach through the darkness, flipping every switch and lever without showing a face.

Vulnerabilities

Trading AI agent lets anonymous callers run shell commands as root

Three GitHub advisories published on 2 October 2026 describe a dozen flaws in Vibe-Trading's LLM agent, including an unauthenticated path to full container takeover.

2 Oct 2026

A figure at a control panel redirects flowing water pipes inward, flooding the building's hidden basement chambers while guards watch helplessly from above.

Vulnerabilities

Trigger.dev webhook alerts can be aimed at internal servers

Two advisories published on 2 October 2026 say any organisation member can make the Trigger.dev control plane POST signed alerts to internal and cloud metadata addresses.

2 Oct 2026