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 2 of 29
340 posts
Nov 12, 2026 · 10 min readCVE-2026-118740: .NET Aspire Dashboard NotFound Route Deletes Telemetry Without AuthMicrosoft published GHSA-4g86-54fm-h9h8 on November 10–12, 2026 for CVE-2026-118740, a missing-authorization flaw in the .NET Aspire Dashboard. Releases through 13.6.4 expose an unauthenticated NotFound handler that can wipe telemetry in AppHost and standalone deployments and, in browser-token mode, expose the browser token. Version 13.6.5 is the first fixed dashboard release; 13.6.3 and 13.6.4 are not sufficient upgrade targets.
Nov 12, 2026 · 11 min readCVE-2026-125926: AWS mcp-gateway-registry SSRF Lets Authenticated Users Probe Internal NetworksOn November 11, 2026, AWS published bulletin 2026-180-AWS for CVE-2026-125926, a server-side request forgery flaw in the open-source mcp-gateway-registry. An authenticated operator who may register MCP servers could set a crafted upstream address so the gateway contacted internal network services and, in some cases, returned those responses. Version 1.35.0 validates upstream targets on new registrations; stored entries still need manual review.
Nov 12, 2026 · 9 min readCVE Volume Hit a Record in 2026: Why It Is Spiking and What to DoA chart of monthly CVE counts going vertical is circulating, and the numbers behind it are real: about 88,700 CVEs through October 29, 2026. But part of the jump is counting rules, only a tiny share is exploited, and NVD enrichment has collapsed. Here is how to read it if you run coding agents.
Nov 12, 2026 · 11 min readDatadog ARTEX AWS Hunting: CloudWatch Secret Search, Cognito Abuse, and Defender PlaybooksDatadog Security Labs published field observations of ARTEX against AWS in November 2026—anonymous S3 reads, CloudWatch log searches for AKIA keys and financial strings, and noisy Cognito self-registration. On November 12, Dark Reading gathered defender quotes that match the same hygiene story. Here is the attack chain from primary sources, what to query in CloudTrail, and what changes for developers whose repos still hold long-lived cloud keys.
Nov 12, 2026 · 11 min readFaceStealer: DOJ Says North Korea Targets Open-Source Maintainers With Deepfake MeetingsThe U.S. Department of Justice announced on November 12, 2026 that it seized FaceStealer infrastructure used in more than $400 million in cryptocurrency theft and described a social-engineering chain that increasingly targets open-source software developers to enable supply-chain compromise. This post walks through the mechanism the DOJ and FBI published, how it overlaps with the TraderTraitor Binance theft response, and what teams running AI coding agents should verify on maintainer laptops.
Nov 12, 2026 · 9 min readGSocket on Linux Servers: Persistent C2 That Survives Your App PatchMicrosoft Security published research on November 11, 2026 showing threat actors planting the open-source GSocket tool on compromised Linux servers. Patching the original web app does not remove the relay client. Telemetry rose 356.2% from March to August 2026, mostly on PHP and Apache estates—and some actors tried CVE-2026-43284 for root.
Nov 12, 2026 · 11 min readMicrosoft MDASH Found 59 Linux Kernel Bugs: 33 CVEs and Eight Root ExploitsOn November 11, 2026 Microsoft published "Lessons from 59 Linux Kernel Bug Reports". Its FORGE Lab used an AI scanning system called MDASH to find bugs up to 21 years old, with working local root exploits for eight CVEs. This post covers the pipeline, the rhashtable exploit chain, and what to patch on machines that run coding agents.
Nov 12, 2026 · 10 min readPrompt Injection as Social Engineering: Oblomovka on Agents, Deputies, and CapabilitiesOn November 11, 2026, Danny O'Brien published that prompt injection is social engineering aimed at silicon—models obey smuggled instructions the way humans obey phony authority. He ties the confused-deputy problem to capability-based security while warning cap-security alone cannot fix approval-clicking. This post translates the essay for Claude Code and MCP operators with concrete checks.
Nov 12, 2026 · 10 min readOpenAI Daybreak Red Team and Defense Factory: Agentic Pen Test and Patch Loops for Coding TeamsDuring OpenAI's November 12, 2026 "Cyber at the Frontier" keynote, the company pitched Daybreak Red Team for authorized continuous pen testing and Daybreak Defense Factory for agent-driven remediation with human-reviewed merge requests. The launch sits beside the open Codex Security CLI and a public reference architecture for continuous defense. Here is what each product does, what data leaves your perimeter, and how it intersects with coding agents on developer laptops.
Nov 12, 2026 · 10 min readOpenAI Fuzzy Robustness: Automated Red Teaming Trains Jailbreak-Resistant Frontier ModelsOn November 12, 2026, OpenAI Alignment published research on fuzzy robustness—how reliably frontier models withhold dual-use help under adversarial pressure in cyber and biology. The team scaled automated jailbreak search with a worst@n success criterion, trained on synthetic robustified prompts, and reports higher robustness on GPT-6 Astra and Sol without extra benign refusals. Coding teams should read this as a model-layer update, not a substitute for watching what agents actually do locally.
Nov 12, 2026 · 11 min readAgentRun: Open-Source Managed Agents, Learning Git Folders, and Sandbox BoundariesAgentRun is an Apache-2.0 control plane that wraps Claude Code, Codex, OpenCode, and pi with durable execution, remote sandboxes, MCP tools, and a learning system that commits notes and skills into per-agent git folders. This post walks through the README security model, approval gates, and what changes for teams that already run local coding agents.
Nov 12, 2026 · 10 min readProject Glasswing at Seven Months: 37,966 Findings, 2 Exploited, Claimed vs VerifiedA VulnCheck review dated November 9, 2026 says Anthropic's Project Glasswing ledger lists 37,966 findings, yet only 442 attributed CVEs exist and just 2 are known exploited. This post separates what is verified from what is claimed, and explains what it means if you run AI coding agents.Page 2 of 29 · 340 posts