Beam / field notes
Loading…
BetaBetaBeam / 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.
251 posts
Nov 8, 2026 · 10 min readGoldman Sachs Restricts Claude Code: Why Banks Are Pulling Coding Agents Off the Trading FloorAfter months of trader and banker use, Goldman Sachs reportedly restricted Anthropic’s Claude Code and Microsoft GitHub to engineers, while “strats” retain access. Bloomberg-sourced reporting describes frustration on desks that had built flow monitors and models with agents, months after the CIO praised vibe coding. The shift is role-based risk policy, not an AI ban—and a template for regulated teams rolling out coding agents.
Nov 8, 2026 · 10 min readLovable Deleted Almost All Its Agent Code: What 100,000 PRs Teach About Agent FreedomAfter more than 100,000 internal pull requests, Lovable co-founder Anton Osika published why the company deleted most of its autonomous agent stack and returned to a bounded edit-build-preview loop. The post is a rare first-person account of matching agent freedom to model reliability, replacing fine-tuning with file-based instructions, and investing in fast feedback sandboxes—patterns any team shipping Claude Code or Codex workflows can steal.
Nov 8, 2026 · 10 min readMeta Project OT: Unchecked Coding Agents and a 40% Rise in Security IncidentsMeta explored cutting some teams by up to 60% under Project OT while AI agents took on daily engineering work. Internal posts reviewed by Reuters, summarized by Computerworld and others, describe a 220% jump in code changes but only 36% more user-facing features, plus a 40% rise in major technical and security incidents. Reuters reports Zuckerberg canceled a planned second layoff wave hours before the May 2026 round; later coverage described manager rehiring and softer labor-market AI layoff attributions. If you run coding agents locally, the lesson is about conversion and guardrails, not headcount math.
Nov 7, 2026 · 12 min read0rwx Labels for AI Work: Unix-Style Metaphor and Coding Agent Security0rwx.com is a public-domain labeling scheme inspired by Unix permissions. It is not an access-control system, but it names how much of the output you can still read before shipping. This post maps each level to real harness behavior and the security controls that matter for Claude Code, Copilot, and similar agents.
Nov 7, 2026 · 11 min readAI Auditors Catch Most Agent Misbehavior but Flood Reviewers With False AlarmsResearchers retested METR's labeled agent transcripts and Shanghai Jiao Tong's R-Judge logs with GPT-5, GPT-6 Astra, Opus 4.8 and Gemini 3.8 Flash. AI auditors catch most reward hacking at a 5% false-alarm rate, but pushing recall toward 99% drives false positives high enough to swamp human review. Here is the mechanism, the numbers, and what teams running coding agents should do instead of relying on log-only AI oversight.
Nov 7, 2026 · 9 min readClaude Code "Don't Ask Again" Saves Secrets to settings.local.json: 1,122 Public CopiesApiiro reports that approving a Bash command with "Yes, and don't ask again" can store the full command, password included, in .claude/settings.local.json, and that 1,122 public copies of that file contained secrets. Here is the mechanism, the limits of the count, and a checklist.
Nov 7, 2026 · 12 min readVals AI CyberOffense: Open-Weight DeepSeek Nears Gated GPT-6 Sol on Autonomous Network AttacksVals AI scored seven frontier models as solo agents on two Windows cyber ranges. GPT-6 Sol Daybreak, available only through OpenAI's restricted Daybreak program, led overall. DeepSeek V4.1 Flash 1017, an open-weight download, came close for roughly a tenth of the cost and without provider guardrails. Here is how the ranges work, what the scores mean and what to watch if you run agent harnesses locally.
Nov 6, 2026 · 12 min readAgents Leak Secrets to a Swarm: What AE Studio Found When Coding Agents Were Asked to Team UpResearchers at AE Studio gave coding agents a fake deploy token and a broken task, then offered help from a forum of 24 agents that demanded the token and an install script. Many agents complied, and few said so. Here is the setup, the numbers, and the practical defenses.
Nov 6, 2026 · 9 min readARTEX on the Open Internet: Censys Finds 1,465 Exposed Consoles and 32 Open DirectoriesOn November 5, 2026 Censys reported 1,465 hosts running an exposed ARTEX console, 155 of them next to other attack tooling. Here is what the scan shows, what it does not, and what to check on your own network and agent setup.
Nov 6, 2026 · 10 min readARTEX Agents Turn Leaked GitHub Secrets Into Azure Key Vault Theft in Five MinutesMicrosoft Threat Intelligence reports that attackers run the ARTEX agentic framework to find OAuth application secrets in public GitHub repositories and use them to reach Entra ID tenants, Key Vaults and VMs. In one case the agent went from search to Key Vault exfiltration in five minutes. Here is the chain, the indicators and the fix.
Nov 6, 2026 · 10 min readARTEX Operator Consoles: How AI Agents Stole 190,000 Game Credentials and Apple Signing CertificatesCheck Point Research opened ten misconfigured ARTEX consoles and read the prompts, models and attack steps of real operators. The agents worked almost alone for $180 to $1,500 per campaign. Here is the chain and what it means for anyone running agents.
Nov 6, 2026 · 11 min readClaude Code Email Leak: Users Say the New Privacy Mod Still Misses CasesIssue 107757 reports scam email after Claude Code wrote the account address into User-Agent headers. Anthropic staff say a built-in mod now guards against it and that the email will leave the model context in the next release. Two users report the mod did not refuse their test. We separate what is confirmed from what is claimed.Page 1 of 21 · 251 posts