Beam / field notes
Loading…
BetaBeam / field notes
Loading…
Beam / field notes
What we're seeing as we watch AI coding agents work — flagged actions, skill/MCP scan findings, and how to keep Claude Code, Codex, and the rest honest without slowing them down.Page 22 of 38
451 posts
Oct 30, 2026 · 10 min readHugging Face trust_remote_code Miner Campaign Returns, and a Fake Solana Kit Targets Coding AgentsTwo separate campaigns show the same weakness: code that runs because something you trust chose it. Pluto Security tracked a Hugging Face miner campaign that came back after takedown, and Air Security described a GitHub repo that ranks first when an agent looks for a Solana starter kit. Here is the mechanism and what to check.
Oct 30, 2026 · 13 min readn0n Ransomware Ran OpenCode With 154 Models and Left Its Workspace Open: What SOCRadar FoundAn extortion crew called n0n used the open-source OpenCode coding agent to write and debug hundreds of attack scripts, then left its Python file server open to the internet. SOCRadar read the 1,719 files. Here is the attack chain, what is and is not confirmed, and what it means for anyone running coding agents.
Oct 30, 2026 · 11 min readNIK C2 Malware Falls Back to Your OPENAI_API_KEY: What Talos Found and What to CheckCisco Talos reported NIK C2, a Go-based Windows implant with an ai_generate command that calls the OpenAI API from the victim machine and, if the operator supplies no key, reads the OPENAI_API_KEY environment variable. This post covers the mechanism, what is and is not confirmed, and how to keep agent API keys out of reach.
Oct 30, 2026 · 9 min readOpen-Weight Model Turns Firefox Patch Gap Into Working Exploits: What Epoch AI FoundEpoch AI reports that an open-weight model, given only public fixes, built working Firefox exploits for about $1,000 per version. The result changes how long you can leave agent tools, browsers and dependencies un-updated. Here is the method, the limits the authors state and what to do.
Oct 30, 2026 · 12 min readHow OpenAI Monitors Frontier Agents for Misalignment: Four Layers and the Recall NumbersOn October 29, 2026 OpenAI described the layered system that watches its models during training, evaluation and internal Codex use. It also published recall figures, including threat types the monitor missed entirely. Here is how the layers work, what the numbers say, and what to borrow if you run coding agents.
Oct 30, 2026 · 7 min readPolymarket Third-Party Script Compromise: A Vendor Script Drained About $2.6 Million From Logged-In UsersOn October 29, 2026 Polymarket put its exchange in cancel-only mode after a compromised third-party provider led to a crypto-draining script running for logged-in users on polymarket.com. Researchers and BleepingComputer tie the script to DebugBear, a performance-monitoring vendor, which Polymarket has not confirmed. About $2.5 to $2.7 million was traced. This post covers the timeline, what is confirmed, and how to defend a site against a vendor script turning hostile.
Oct 30, 2026 · 10 min readUnauthorized Codex and Claude Code Relays: Sandbox Off, Traffic ReroutedSymantec's Threat Hunter Team describes a reseller that installs genuine OpenAI Codex and Anthropic Claude Code, rewrites their config to a relay, sets full-access sandbox mode and turns telemetry off. Here is the mechanism, the config keys to check and what to do.
Oct 30, 2026 · 10 min readVAPOR PANDA and ARTEX: How One AI Agent Controller Targeted 200+ Organizations in 11 DaysCrowdStrike OverWatch says a China-nexus group ran the ARTEX agent framework from one controller against more than 200 organizations in 14+ countries between October 17 and 28, 2026. Nine were confirmed compromised. This post separates targeting from confirmed breaches, walks through the chain, and lists what to hunt for.
Oct 29, 2026 · 10 min readContractor's Claude Code Agent on a DOD Network: Claimed vs VerifiedAnonymous sources told WIRED that a contractor connected a personal Claude Code account to a DOD system, and the agent sent credentials and a network map to a third-party site. Accenture and the DOD dispute or decline to confirm. Here is the claimed-versus-verified breakdown and the checks it suggests for your own setup.
Oct 29, 2026 · 6 min readAU Small Finance Bank Breach Claim: ScarlettGroup, KYC and Aadhaar Data, Claimed vs VerifiedOn October 29, 2026, Hackmanac reported that a group calling itself ScarlettGroup claims to have breached AU Small Finance Bank and listed KYC, Aadhaar, PAN, credit and loan records on a leak page with a negotiation countdown. The bank has only asked people to contact it by direct message. This post separates the claim from the evidence.
Oct 29, 2026 · 10 min read26,400 Live Cloud and AI Keys Found in Public Gitee Repos: What Coding-Agent Users Should CheckIndependent researcher Ben Zimmermann scanned 7,256,531 public Gitee repositories and found 26,400 credentials that still authenticated, including thousands of AI provider keys. This post covers the method, the numbers, the limits of the claim, and what it means if a coding agent commits code for you.
Oct 29, 2026 · 8 min readLinux CVE-2026-98432: TCP Use-After-Free Patch Status and the "MPTCP Only" Claim CheckedCVE-2026-98432 went viral on October 29 as a 9.8 kernel bug that only matters if you run MPTCP. The CVE text describes the ordinary IPv4 TCP handshake path. Here is what is fixed, what is still pending, and what we could not confirm.Page 22 of 38 · 451 posts