Corelight: MCP servers can pull unreviewed packages with admin access
On one lab workstation, 19 rebuilds of one pinned MCP release produced 15 environments as transitive PyPI packages drifted.
Stories per day · 30 days
44 RSS
Relatedanthropic27openai27prompt-injection24agent-security19claude19mcp19credentials13microsoft12nvidia12identity9meta9benchmark8credential-theft8open-source8ai-governance6cairn6cloudflare6muse6
909 stories, newest first · one card per story
Stories tagged #ai-agents
Databricks updates Genie to run multi-step jobs unsupervised
CEO Ali Ghodsi demoed it on surveys and a failed sales pitch; Unity Gateway can cap usage and swap models.
Plurai raises $13 million to simulate and tune AI agents
The seed round, led by Team8, funds digital twins that post-train small open-weight models for customer workflows.
Microsoft shows weight edits can hide backdoors in AI agents
Changing fewer than 10 dormant neurons can make a triggered agent run attacker code; Microsoft has not seen this used.
Neumann says GenDB-V benchmark gains come from query overfitting
Reported 22× to 61× speedups over Umbra and DuckDB rely on precomputed structures and hard-coded TPC-H layout.
Token says browser agent Rehearsal broke its own identity rules
The vendor says apps cannot see an agent inside a user session and lists controls to use before standards land.
Bay raises $50 million to control enterprise AI agents
Founded in February 2026 by three Unit 8200 alumni, Bay inventories agents and can deny actions in milliseconds.
Orca adds discovery for shadow AI apps and agents
Agentless scanning looks for builder fingerprints and maps agent permissions, credentials and access to sensitive cloud data.
Obsidian launches runtime policy engine for AI agent actions
Policies can allow, block, or require approval, including for OpenAI Codex and MCP tool calls.
Threat actors adopt open-source Artex AI pentest framework
The creator deleted the GitHub repository in late October after abuse, but forks still support credential and cloud attacks.
Near AI breach exposes about 100 users and data from 18 agents
The company says an attacker used a remote code execution flaw in SGLang to reach its inference infrastructure.
Google releases Gemini 4 Carbon with agent swarm feature
Paying developers and AI Ultra users get Carbon today at $2 per million input tokens, about a fifth of GPT-6 Astra.
Claude Code permission file exposes secrets in 1,122 public GitHub copies
Apiiro found 1,379 secrets, including database passwords and AWS keys; one notified company confirmed a live key.
4
Developers share ways to control quality of AI-written code
The original poster says Claude-generated features lack a shared behavior model, and commenters suggest tests, docs, and build-failing rules.
eBay blocks third-party AI agents, including Meta's Muse, from browsing
Muse could pull eBay listings until this weekend; eBay already bans buy-for-me agents and says some agent access is restricted.
Wiz relaunches Cyber Model Arena to score agent cyber tasks
Version 3.9 tests 59 agents on 327 challenges with a 1,000-second budget and deterministic grading.
NVIDIA donates OpenShell to CNCF and adds Kubernetes AI tools
A Crusoe test of Snapshot cut GLM-5.3 readiness from about eight minutes to 31 seconds.
Meta launches Spaces for project work with Muse agents
Spaces sit in the Muse sidebar and can stay updated from connected third-party apps without a new prompt.
WhiteMatter raises $20 million Series A for AI agent security
Blumberg Capital led the round, bringing total funding to $24 million for the former QuilrAI.
Adobe previews local MCP agents and always-on Firefly automation
OpenAI, Google, and Anthropic agents could edit local Creative Cloud files; a private beta is set for early 2027.
1
Cyanogen co-founder unveils Kora, an AI desktop OS for 2027
The free Linux system targets Intel, AMD, and Snapdragon PCs, with enterprise Humain OS also planned for 2027.
Survey: security teams plan most monitoring investment for AI agents
Nearly half of respondents said their organizations cannot effectively baseline normal AI agent behavior.
1
Paper proves no finite rulebook can certify every safe program
Watching can catch every prohibition violation, but no safe program can be cleared for good.
Anecdotes CEO says GRC managers must learn to govern agents
Kuznitsov argues agents already do control testing, evidence collection and framework mapping, and names three skills.