AI Builders Digest

Anthropic details the security boundaries around Claude agents

Anthropic says containment matters more as agents gain access to files, networks, and internal services. Its Claude Code sandbox cut permission prompts by 84%, but a phishing test still led Claude to send credentials out in 24 of 25 tries. A separate Cowork incident showed that even an approved API domain could carry files to an attacker’s account, prompting Anthropic to add a proxy that checks which credentials a request uses.

• #agents
• #security

OpenAI says affected OpenCode users are back in business

Thibault Sottiaux says OpenAI worked with OpenCode to restore access for users who are also OpenAI API customers and refund credits consumed in an attack. He also says defenders can request Daybreak Blue access for a cyber permissive setup before an incident occurs.

• #security
• #products

Anthropic traces Claude Code quality complaints to three changes

Anthropic says a lower default reasoning effort, a bug that repeatedly cleared prior reasoning, and a brevity instruction each hurt Claude Code users in different ways. All three were resolved by April 20; the API and inference layer were unaffected. The company says it is broadening evaluations for prompt changes and resetting subscriber usage limits.

• #products
• #evals

Reflection AI’s Beam pairs open weights with large scale reinforcement learning

Reflection AI released Beam, an open model with 500 billion total parameters and 23 billion active at a time. Misha Laskin says Beam is three to four times more reasoning efficient than models in its capability class, a result he credits in part to a large reinforcement learning run. He argues that open models also give more defenders access to capable tools: “with enough eyeballs, most security and safety vulnerabilities become shallow as well.”

• #open-source
• #models
• #security

Aaron Levie sees agent swarms turning research into a compute problem

Levie argues that agent swarms could parallelize work now limited by people’s ability to plan, execute, and review it. He points to cybersecurity, life sciences, materials research, and QA as fields where coordinated agents could make available compute a major constraint, if models become capable enough.

• #agents
• #research

Swyx says demand is shifting toward deep data engineering

Swyx sees a widening gap between people who can assemble a retrieval app and engineers who understand data intensive systems. He says demand for the latter is stronger than at any point in his tech career and is recruiting for that skill set.

• #hiring
• #data

Guillermo Rauch adds game features with a low cost model run

Rauch says he used DeepSeek V4.1 Flash through AI Gateway to add sound, improve collision detection and a ramp, and make a camera follow a car. He estimates the run cost about 30 cents.

• #models
• #products

Peter Yang argues for work trials over interviews

Yang says hiring teams should make short work trials the norm. He shares an example from Anthony Pompliano in which a candidate built a requested feature in two weeks, giving the team evidence of how the person worked while shipping the feature.

• #hiring

Thariq explores games where building characters is the main draw

Thariq says he sometimes enjoys coding small parts of a game more than playing it. He is considering an open source game where players create their own characters with AI and the game balances them, though he has not settled on the play mechanics.

• #games
• #open-source