Agents: Sandboxes, transparency, governance, and DSLs

How I tricked Claude into leaking your deepest, darkest secrets. A researcher exploits Claude’s web_fetch linking rules to exfiltrate private memories, forcing Anthropic to remove link-following in fetched content. Outcome engineers must treat tool and memory APIs as attack surfaces—design stricter sandboxing, provenance, and least-privilege tool interfaces to prevent silent data leaks (Principles 07,14).

Perplexity launches SPACE sandbox to make its AI agents secure and powerful. Perplexity ships SPACE, a secure sandbox that lets Computer agents operate at full capability while isolating risk. Adopt sandbox patterns like SPACE to balance agent capability and containment, embedding runtime isolation, capability diplomacy, and policy enforcement into your agent runtime (Principles 07,14).

Codex Multi-Agent V2 update raises developer concerns over agent transparency. OpenAI’s Codex update hides inter-agent instructions, prompting developer demands for auditable instruction traces and clearer provenance. If you build multi-agent systems, require transparent instruction logs, deterministic handoffs, and audit hooks so agents are verifiable components in your delivery pipeline (Principles 13,15).

Port releases Port AI Builder vibe coding platform for dev and platform teams. Port launches an agentic SDLC platform that lets teams build governed workflows in natural language with built-in human approvals. Treat agentic platforms as orchestration hubs: define human-in-the-loop gates, policy-as-code, and artifact provenance from design through deployment (Principles 09,06,15).

DSLs Enable Reliable Use of LLMs. Martin Fowler argues DSLs provide LLMs with precise, executable domain models and make generated systems reliable by making the DSL the source of truth. Outcome engineers should invest in domain-specific languages and executable contracts to constrain model behavior, reduce context drift, and make agent outputs testable and auditable (Principles 06,13).