Synthesis history

The security vulnerability picture has expanded from the single previously-tracked CVE (CVE-2026-25725) to a documented multi-vector pattern: John Stawinski's February 2026 prompt injection to RCE [^19799], Check Point …

CVE-2026-25725 — a Claude Code sandbox escape via persistent configuration injection in settings.json, patched quietly by Anthropic [^22296][^23129] — converts the prior synthesis's security architecture argument from t…

Three developments materially extend the prior synthesis. First, Anthropic's containment post-mortem [^21652] adds security architecture as a new analytical layer — finding that 93% human approval fatigue and probabilis…

Two meaningful developments this pass. First, the code-as-agent-harness thesis advanced from a survey-level prescription [^20654] to a dedicated academic formalization (arXiv 2605.18747) [^21363] that defines three spec…

The primary new development is a Meta+Stanford+Illinois survey paper arguing that code, not natural language, should be agents' primary working layer — directly addressing the state-loss and error-hiding failures of tex…

Three developments distinguish this pass. First, a Meta paper showing coding agents improve significantly with structured summaries of prior attempts over raw logs [^12850] adds a constructive engineering response to th…

Three developments distinguish this pass. First, Anthropic's engineering blog post 'Effective harnesses for long-running agents' [19723] is now tracked as a primary source rather than through InfoQ intermediary coverage…

This pass adds three significant developments. First, the 'Towards a Science of AI Agent Reliability' paper has achieved full academic institutionalization: a Princeton CITP seminar, a PREreview peer review, and an acti…

This pass adds two significant developments not present in the prior synthesis. First, a wave of practitioner and commentator voices (May 18–23) has coalesced explicitly around 'reliability' as the meta-frame for produc…

A cluster of research findings published in May 2026 collectively challenge three dominant assumptions in AI agent design. • A Stanford study argues that under equal computational budgets, a single LLM outperforms coord…