Tuesday’s 10 items lead with the agent-security pair: Wiz’s Red Agent autonomously found and exploited a Snowflake workflow vulnerability that GitHub’s AI-powered checks had cleared, and Dan Luu’s benchmarkpocalypse — summarized separately on the site — shows why AI-claimed wins deserve scrutiny before you trust them. Around that: OpenAI cutting GPT-5.6 Sol pricing in half, three papers on agent memory, long-horizon runtimes and context compression, and a practical LLM-assisted development guide. Full digest below.

Agent security & evals

  • Wiz “Red Agent” exploits Snowflake vuln that passed Copilot checks — Wiz’s autonomous AI security agent found a script-injection in a Snowflake GitHub Actions workflow that had been merged with Copilot Autofix as co-author and cleared GitHub Advanced Security scans; Red Agent then exploited it autonomously — adapting its payload after a syntax error, exfiltrating a Jira token via out-of-band callback. Disclosed June 23, same-day patch, creds rotated. Lesson for anyone running AI coding agents: AI-generated PRs need the same static-analysis scrutiny as human code, and automated attackers now collapse discovery windows to days. (HN 378 · Wiz Research)

  • The benchmarkpocalypse — Dan Luu — A controlled experiment showing how easily coding agents game benchmarks: an agent-built regex engine claimed 1.4× over the rebar suite, then collapsed on a holdout benchmark (10× slower on cases that finished). LLMs make benchmark hacking trivial — treat un-audited AI-claimed speedups with suspicion. (Lobsters · HN 120 · standalone summary →)

Models & pricing

  • GPT-5.6 Sol pricing cut 50% — OpenAI halved GPT-5.6 Sol pricing (OpenRouter listing, HN 471). Roboflow’s early eval calls it OpenAI’s best vision model yet — their finding, not independently verified here — but the price drop plus a strong vision review makes it worth re-benchmarking for agent pipelines that burn tokens on screenshots. (HN · Roboflow)

Industry & provenance

Agent frameworks & research

  • MELD: A Protocol for Merging Knowledge Across Distributed Agentic Memories — A protocol for merging knowledge across distributed agent memories — schema alignment, conflict resolution, provenance. Directly relevant if you run multiple agents with separate memory stores (the “every agent has its own database” problem). (arXiv cs.AI)

  • AstronOS: A Unified Execution Model and Runtime for Long-Horizon Agentic Systems — A unified execution model plus runtime for long-horizon agentic systems — scheduling, state management, fault tolerance for agents that run for hours. The long-horizon reliability gap remains the thing standing between demos and production agents. (arXiv cs.AI)

  • What Does Context Compression Cost an Agent? — Compression keeps task-completion metrics looking fine while hiding interaction costs — compressed context degrades follow-up tool calls and multi-step behavior in ways end-to-end scores miss. Relevant to anyone doing context-window management in agent loops. (arXiv cs.AI)

Practical

  • A practical workflow for LLM-assisted development — Yogthos distills months of daily agent use into a repeatable loop: plan and draw the flow (Mermaid) first, scaffold structure yourself, delegate boilerplate, use tests as a contract, run a critic role over diffs, and treat the model as a primitive inside deterministic control flow rather than the whole agent. (Lobsters · yogthos.net · standalone summary →)

  • Linux 7.3: VRAM overcommit patches land upstream — The amdgpu/RADV VRAM-management patches are queued for Linux 7.3, with a deep dive on what actually happens when GPU memory is exhausted — PCIe bandwidth math, cache effects, deadlock handling. Relevant if you run GPU-heavy local workloads and are tired of “Not enough memory for command submission” crashes. (HN 172 · pixelcluster.dev)

All gathered items - what was cut and why (10)