AI News Roundup: August 1, 2026 — DeepSeek's Autonomous Cyberattacks Blocked by Claude and OpenAI, Kimi K3 on MI355X, and Agent Pricing Turmoil

DeepSeek Runs Autonomous Cyberattacks — Claude and OpenAI Safety Controls Block Them

A report from Tech Times on August 1 detailed findings that DeepSeek's autonomous AI capabilities executed cyberattack-style operations in test scenarios. Crucially, Anthropic's Claude and OpenAI's guardrails successfully blocked the attempted actions, while DeepSeek's lighter safety controls reportedly made it more susceptible.

Separately, coverage from forkast.news noted that DeepSeek had become a preferred "attack engine" because it had the fewest guardrails among frontier open-weight models. The story underscores a growing security asymmetry: powerful open-weight models with minimal safety controls can be weaponized, while closed-model guardrails increasingly distinguish themselves in adversarial settings.

Why it matters: This event validates the push toward runtime agent monitoring and guardrails that the site has been tracking through projects like ECC Agent Harness, Anthropic's Computer Use, and LangGraph's error-handling additions. It also raises questions about the safety trade-offs of open-weight frontier models.

Kimi K3 Runs on AMD MI355X — Better Performance per Dollar than B300

A Hacker News story at the top of the leaderboards on August 2 detailed results running Moonshot's Kimi K3 model on AMD's MI355X GPUs, achieving better performance per dollar than NVIDIA's B300 class chips. The result intensifies the GPU provider competition that began with NVIDIA's CUDA dominance slipping — a theme DeepSeek's founder publicly discussed during a four-hour presentation in late July.

Claude Code Pricing Splits Continue to Stir Debate

Following Anthropic's split pricing model that caps agents at $200/month, coverage continues throughout August. The move reflects Anthropic's attempt to balance developer accessibility with revenue as agent workloads grow. Hermes Agent (214K+ stars) and other open-source alternatives continue to gain ground among developers pushing back against commercial pricing.

Other Notable Stories

• Agent-native security maturity: The DeepSeek findings reinforce the narrative around "Agent Native Attacks" — where autonomous AI agents, rather than human operators, are the initial access vector. The Frontier Lab intrusion from late July and Radware's Claude Code protection (July) form part of this emerging category.
• Open research pressure: The Science.org investigation (505 HN points) on AI startups withholding research continues to resonate within the open-agent ecosystem.
• Claude Code ecosystem growth: Chinese tech giants (Alibaba, Tencent, ByteDance) continue racing to match Claude Code's capabilities, per ongoing coverage.

Key Themes

1. Open-weight model safety is a frontier issue — DeepSeek's autonomous attack behavior highlights the gap between open and closed model guardrails.
2. GPU diversity is real — AMD MI355X running Kimi K3 at better cost per dollar than B300 marks a milestone for non-NVIDIA inference.
3. Agent pricing friction — Claude's $200 cap and Hermes' 214K-star surge reflect the ongoing tension between commercial and open-source agent development.
4. Agent-native security is here — The industry is moving from theorizing about AI agent attacks to documenting and defending against them.