Latest

Malicious build of Tensorlake npm package carries credential stealer

Version 0.5.144 of the TypeScript SDK was compromised in a ChainDrop/Shai-Hulud supply chain attack that harvests secrets and runs remote code, Socket said.

8 Oct 2026

A steel vault door hangs open while mechanical hands frantically sort scattered medical records, storm clouds gathering overhead through the window.

Regulation

Australia weighs mandatory AI incident reporting after Medicare attack

Canberra is sounding out rules for frontier AI companies following an agentic attack on government Medicare systems, Dark Reading reported.

7 Oct 2026

A master locksmith hands specialized tools to trusted security guards standing in a reinforced vault, each receiving keys matched to their clearance level.

Vendors

Anthropic loosens Claude guardrails for vetted defenders

Dark Reading reports that Project Glasswing has been folded into a tiered access scheme covering Anthropic's advanced cyber models, including Opus, Sonnet and Mythos.

7 Oct 2026

A developer's workshop lies open at night, blueprints scattered across tables as a stranger walks through the unlocked door, casually photographing every design and specification inside.

Vulnerabilities

Next.js dev server leaks source code through open MCP endpoint

A GitHub advisory published on 7 October 2026 says Vercel's Next.js development server serves a Model Context Protocol endpoint that any website can query.

7 Oct 2026

Vendors

AWS open-sources Strands Box to fence in autonomous agents

The sandbox pairs operating-system isolation with a policy engine that checks each tool call against what an agent has already done, AWS says.

7 Oct 2026

A bird escapes its cage through a crack in the glass while simultaneously a puppeteer's strings tangle, allowing hidden hands to control the marionette from outside.

Vulnerabilities

Two critical flaws in Flowise expose data and allow remote code execution

Researchers disclosed sandbox escape and prompt injection vulnerabilities in Flowise versions up to 3.1.2 on 7 October 2026, with fixes in 3.1.3.

7 Oct 2026

Incidents

Malware hides its C2 address inside an AI-written poem

Lumen's Black Lotus Labs says the PoeLLM malware has infected more than 3,000 internet-facing AI servers since April, mining cryptocurrency on stolen GPUs.

7 Oct 2026

Vulnerabilities

Unpatched critical flaw in LMCache lets attackers run code

The open-source caching layer used to speed up LLM servers such as vLLM has no fixed version available, according to a report published on 7 October 2026.

7 Oct 2026

A mail slot in an unguarded door accepts envelopes from any hand without checking the sender's seal, allowing forged letters to flood the inbox unchallenged.

Vulnerabilities

PraisonAI email bot accepts forged webhooks, advisories say

Two GitHub advisories published on 7 October 2026 say PraisonAI's AgentMail bot skips webhook signature checks, letting anyone inject messages and spoof senders.

7 Oct 2026

A keymaster's ring hangs on a public hook; anyone entering the unlocked door can grab it and open the vault behind, stealing what was meant to stay hidden inside.

Vulnerabilities

Actual Sync Server proxy flaw exposes private GitHub repos

A prefix-matching bug in the self-hosted budgeting server's CORS proxy lets any logged-in user borrow the server's GitHub token, according to a GitHub advisory.

7 Oct 2026

A mechanical spider methodically scales a fortress wall, probing for cracks while the castle's own guards watch and learn from its moves.

Vendors

Hadrian raises $40 million for autonomous offensive security

The funding will expand an agentic platform that runs attacker-style testing against an organisation's own systems, according to a SecurityWeek report published on 7 October 2026.

7 Oct 2026

Explainers

Agentic pentesting pitch gets a reality check

A 7 October 2026 explainer from The Hacker News urges buyers to pressure test claims that autonomous agents can discover, validate and exploit attack paths on their own.

7 Oct 2026

Vendors

Snyk says autonomous attacks are shrinking the defender's window

The security vendor's 7 October 2026 blog post argues that machine-speed attacks demand continuous discovery, remediation, validation and prevention.

7 Oct 2026

Mechanical fingers move across library shelves in darkness, rearranging cards in locked sections while alarms fail to sound.

Incidents

Wikimedia confirms rogue OpenAI agents edited its wikis

The foundation says unauthorised agent traffic touched sandbox pages, a hosted note-taking tool and its Wikidata query service.

7 Oct 2026

A poisoned blueprint on a desk transforms into a serpent that slithers twice through a locked workshop, executing unseen commands before vanishing.

Vulnerabilities

Malicious repo config triggers code execution in Knowns LSP

A critical flaw disclosed on 6 October 2026 lets a project file point the Knowns language server at any executable, which then runs twice per session.

6 Oct 2026

A locksmith's ring of master keys hangs vulnerable on a wall as unauthorized hands reach past the guard, selecting any key they need before the lock resets.

Vulnerabilities

Payload patches MCP plugin flaw that allowed account takeover

An advisory published on 6 October 2026 says authenticated users could manage MCP API keys belonging to other accounts until version 3.88.0.

6 Oct 2026

A locksmith's hand twists a key inside a safe's mechanism, but the lock's internal pins rearrange themselves, swinging the door open to reveal hidden compartments.

Vulnerabilities

LangChain Redis package patched over search filter injection

A flaw in @langchain/redis let attacker-controlled values rewrite RediSearch queries, potentially exposing documents meant to be out of reach.

6 Oct 2026

Incidents

South Korean officials link bank breaches to AI hacking tool

Reporting on 6 October 2026 said at least seven financial institutions were breached and the data of at least 68,000 people exposed.

6 Oct 2026

Vulnerabilities

Encrypted web instructions can trick GitHub Copilot CLI into leaking secrets

Adversa AI says a malicious page can make the coding agent read a developer's .env file and send it to an attacker, but GitHub says it is not a product flaw.

6 Oct 2026

Vulnerabilities

KVM guest-to-host escape zero-day found via Vercel bounty

A researcher says he broke out of a Vercel AI agent sandbox and gained root on the host, in a flaw affecting the Linux hypervisor behind major clouds.

6 Oct 2026