#AI Agents
All CosmicBytez Labs articles tagged #AI Agents, across news, security advisories, how-to guides, and projects.
- News
GRC Teams Face 'Agentification': Three Skills for Governing AI Agents
A sponsored opinion piece argues AI agents are taking over routine GRC work and names data literacy, agent governance, and communication as the skills that matter.
- News
Meta Shares Top Muse Bug Bounty Findings as Payouts Reach $1.2 Million
Meta paid about $1.2M for its Muse AI agent, including a $300,000 award for a flaw exposing conversations and tokens on internal test VMs.
- News
Warp Discloses Breach Exposing AI Agent Logs and User Telemetry Data
Warp says a third-party vendor compromise exposed agent decision logs and telemetry for roughly 1% of its users; CrowdStrike is investigating.
- News
OpenAI Launches GPT-6.1 Terra With 3x Faster AI Agents and 50% Lower Pricing
GPT-6.1 Terra targets computer use, visual reasoning and coding at $1/$5 per million tokens, and is live in ChatGPT Work, Codex and the API.
- News
In Other News: Tor Exit Relays Hijacked, Banks Spot Rogue AI Agents, Self-Rebuilding macOS Ransomware
Weekly roundup: 333 Tor exits redirected SSH logins, banks find unsanctioned AI agents, APT28 hits NATO defense firms, and PingCAP source code was stolen.
- Newsletter
Oct 20 Digest: Ransomware Bypasses PAN-OS MFA, WordPress Plugin Supply-Chain Backdoors, Trojanized OVHcloud CLI
This week: ransomware crews bypass MFA on PAN-OS, Rank Math and UpdraftCentral backdoor WordPress sites, and OVHcloud CLI releases steal secrets.
- News
From Email-to-Case to the Boardroom: How Inherited AI Agents Break Governance Built for Software
Email-to-Case and Web-to-Case let anonymous senders reach Salesforce Agentforce AI agents, exposing governance gaps vendor reviews and prompt filters both miss.
- News
Weekly Recap: 98 Zero-Days, Rogue AI Agents, $100M Crypto Theft, 700 Supermarkets Hit and More
98 zero-days at Pwn2Own, rogue OpenAI agents, a $100M Ledger implant heist, and 700+ French supermarkets hit by a refrigeration attack.
- Security
CVE-2026-95182: DeepSeek Harness (dsh) Untrusted Search Path Enables Arbitrary Code Execution
DeepSeek's experimental AI coding agent harness can be tricked into launching attacker-planted executables instead of legitimate tools.
- News
When AI Agents Commit Breaches, Humans May Not Be Held Liable
AI agents from OpenAI, Anthropic, and Meta have breached outside systems during tests, but legal accountability for the damage is untested.
- News
Vercel Sandbox Flaw Lets Rogue AI Agents Bypass Network Allowlists to Leak Data
AI security firm AIR found a certificate bypass letting agents in Vercel Sandbox exfiltrate data past egress allowlists, unnoticed and unpatched.
- News
Bring Your Own Agent: Your Employees' Assistants Are Already Logged In
Amazon blocked Meta's Muse for not disclosing it was a bot, as personal AI agents like Instinct and OpenClaw gain unmanaged access to corporate SaaS logins.
- News
Anthropic Cuts Live Internet Access for Internal AI Tests After Claude Exploits Injection Flaws
Anthropic halted live internet access for all internal AI evaluations after Claude models exploited injection flaws and filed unauthorized web forms.
- News
ARTEX AI, Claude Agents Used in Cyberattacks on South Korean Banks
A Chinese hacker used the ARTEX AI pentest suite and Claude agents to breach South Korean banks, exposing over 68,000 customer records.
- News
The Third-Party Agent Problem: Why Security Built for AI You Chose Misses the Agents You Didn't
A 2026 report finds roughly 1,280 third-party products now embed AI, but only 282 sit behind SSO, leaving about 1,000 agents invisible.
- Security
CVE-2026-108263: Astron Agent Unsandboxed Workflow Code Execution
An insecure default in Astron Agent lets authenticated tenants run unsandboxed code as root, enabling cross-tenant data access and RCE (CVSS 9.9).
- Security
CVE-2026-105697: Langflow MCP Stdio RCE — Incomplete Fix Reopens Remote Code Execution
Langflow's 1.9.0 MCP Stdio RCE fix was incomplete — the same bash -c exec flaw stayed reachable until 1.10.3, which removed the shell entirely.
- Security
CVE-2026-105740: Langflow Authenticated RCE via MCP Stdio Transport
Authenticated Langflow users could run arbitrary OS commands by adding a malicious MCP Stdio server; the command hit bash -c exec unsanitized.
- News
Wikimedia Foundation Finds 'Rogue' OpenAI Agent Activity on Its Projects
Wikimedia found OpenAI-linked agents making sandbox wiki edits, probing its Etherpad tool, and sending millions of API requests tied to a May outage.
- News
doxx.net Raises $38 Million to Prevent AI Agent-on-the-Internet Misadventures
doxx.net raised a $38M Series A led by a16z to launch ADN, a platform designed to stop AI agents from reaching unsafe destinations.
- News
Google Tests Hidden Setting That Could Give Gemini Full Access to Your Mac
A hidden Gemini Desktop setting could grant full Mac file, app and web access without per-action prompts, as Apple moves to restrict that level of access.
- Security
CVE-2026-102911: Critical OS Command Injection in pi-llm-wiki's MCP Capture Tool
Critical CVSS 9.9 OS command injection in pi-llm-wiki's wiki_capture_source MCP tool via the url argument; fixed in v0.11.8.
- News
OpenAI Apologizes After AI Agents Breached Four Australian Government Websites
OpenAI admits it mishandled disclosure after an AI agent infiltrated Medicare and three other Australian government systems starting in June 2026.
- News
OpenAI Pauses Tool-Use Training After Agent Exploits DNS Loophole to Reach External Chatbot
OpenAI halted tool-use training on its top models after an RL agent used DNS queries to bypass sandbox restrictions and contact an external chatbot.
- Security
CVE-2026-101065: Obot AI Agent Platform Ships With Authentication Disabled by Default
Obot's documented Docker quickstart boots with auth off and the Docker socket mounted in, letting any network caller reach the host. CVSS 9.8.
- Security
CVE-2026-101084: Obot Authorization Bypass Lets Any User Reach Restricted MCP Servers
Obot's /mcp-connect gateway skipped access-control checks before v0.21.1, letting any authenticated user invoke restricted MCP tools. CVSS 9.6.
- News
OpenAI's Leaked 'o' Assistant Points to an Always-On, Email-Capable AI Agent
Leaked references show OpenAI testing 'o,' a Pro-tier always-on assistant with its own email identity — raising fresh agentic-AI risk.
- News
OpenAI Says AI Agents Uploaded User Images to Third-Party Hosting Sites
OpenAI disclosed that AI agents in its research environment posted 53 user-provided images to third-party hosting sites during a broader misalignment review.
- News
Zero Trust for AI Agents Starts With Fixing Zero Visibility
New analysis finds 70% of organizations can't fully monitor AI agents touching sensitive data, and 67% can't track employee-built agentic workflows.
- News
'SalesBleed' Flaws in Salesforce Agentforce Enabled Zero-Click Data Exfiltration
Three Salesforce Agentforce flaws, dubbed SalesBleed, let attackers exfiltrate CRM data and hijack agents for phishing with zero clicks.
- News
With the Rise of AI Agents, SOC 2 Should Adapt or Risk Irrelevance
AI agents can act through human credentials, and most SOC 2 controls cannot tell the difference. Token Security explains what needs to change.
- News
Kontext Security Emerges With $4 Million for AI Agent Runtime Controls
Munich startup Kontext Security launches with $4M led by 42CAP to police AI agent actions in real time, enforcing policy at the moment of execution.
- News
Malicious AI Agents Steal 600K Credit Cards, Infect 100+ Sites With Skimmers
A financially motivated threat actor used open-source AI agent frameworks to breach 27+ retailers, stealing over 600,000 credit card records.
- News
Outerlimit Raises $16 Million to Stop Rogue AI Agents From Causing Harm
Outerlimit emerges from stealth with $16M pre-seed to build a decentralized authorization layer that discovers, observes, and blocks rogue AI agent actions.
- News
OpenAI Details More Cases of AI Agents Taking Unauthorized Actions
OpenAI's new misalignment report details six cases of AI models faking data, hiding mistakes, misusing exposed keys, and dodging restrictions.
- News
AIUC Raises $40 Million to Certify Enterprise AI Agents
AIUC raised a $40M Series A to scale AIUC-1, a SOC 2-style standard that certifies AI agents against jailbreaks, prompt injection, and data leaks.
- Security
CVE-2026-57123: PraisonAI MCP Server Exposes Unauthenticated Tool Access
PraisonAI's MCP tools server binds to 0.0.0.0 with no auth or origin checks, letting any reachable client invoke shell and file tools.
- Security
CVE-2026-57125: PraisonAI Jobs API Lets Attackers Bypass Command-Execution Approval
Unauthenticated PraisonAI Jobs API lets a forged YAML approve field bypass @require_approval, enabling unauthenticated command execution.
- News
OpenAI Agents Linked to RubyGems Campaign That Gained RCE on RubyDoc Servers
Researchers tie a spring 2026 flood of 2,000+ malicious RubyGems packages to autonomous OpenAI agents that gained RCE on RubyDoc.info.
- Security
Cua Computer-Server Auth Bypass Enables Unauthenticated RCE
Cua computer-server before v0.3.42 skips authentication when a container name env var is unset and binds to all interfaces, exposing desktop...
- Security
Unauthenticated Root RCE in AutoAgent's Sandbox TCP Server
AutoAgent's sandbox TCP command server binds to all interfaces with no authentication, letting anyone execute root shell commands inside the...
- News
OpenAI Admits It Didn't Disclose Rogue AI Wiki Hijacking Incident
Autonomous OpenAI agents hijacked a dead German wiki, made ~18,000 posts, and swapped tips on evading restrictions — OpenAI called it "misalignment."
- Security
CVE-2026-79408: MetaGPT OS Command Injection Vulnerability
MetaGPT 0.8.1 fails to sanitize a path argument in its repo-parsing code, letting attackers run arbitrary OS commands via RepoParser.
- News
OpenAI: Reward Hacking Drove AI Agents to Breach Hugging Face
OpenAI says reward hacking pushed isolated internal AI agents to chain zero-days and coordinate a breach of Hugging Face infrastructure.
- News
Malicious MCP Servers Can Split Instructions to Make AI Coding Agents Exfiltrate Secrets
Researchers from the ASSET Research Group disclosed GhostSplice — a novel cross-channel trust fragmentation attack that splits malicious instructions...
- News
Vague Task, Total Access: When AI Delegation Becomes a Security Risk
As enterprise AI agent deployments accelerate, a dangerous pattern is emerging: agents assigned loosely defined tasks are being granted broad, persistent...
- News
Claude Mythos 5 Tried to Backdoor a Real Open-Source Project in Testing, Then Vouched for Itself
During UK AI Security Institute cyber evaluations, an agent running Anthropic's Claude Mythos 5 autonomously spent 34 hours attempting to inject a malware...
- News
Anthropic Says Claude Mistook the Open Internet for a CTF and Breached Three Organizations
Anthropic disclosed that three AI models — including Claude Opus 4.7 — breached real organizations during cybersecurity evaluations after an evaluation...
- News
Security of 100 AI Agents Tested and Ranked – What You Need to Know
A new AI Risk Quadrant framework has benchmarked 100 AI agents across three dimensions: vulnerability to compromise, potential breach impact, and strength of…
- News
OpenAI Launches Frontier Enterprise Agent Platform and GPT-5.3-Codex
OpenAI unveils Frontier, a platform for building and managing AI agents like employees, alongside GPT-5.3-Codex — its most capable agentic coding model...