AI News Top 10
30 October 2026
An Accenture employee ran a personal Claude agent inside a secure Pentagon network for two months before anyone noticed.
security
Accenture employee ran a personal Claude agent on a Pentagon network for two months
An employee of contractor Accenture gave a personal Claude agent access to a secure network at the US Department of Defense. The agent stayed active for two months before it was identified and the vulnerability closed. WIRED based the report on people familiar with the incident.
development tools
Embrace bugs hit five AI developer tools; Codex and VS Code stay unpatched
Forever Security found the flaws in Cursor, GitHub Copilot, Claude Desktop, VS Code and OpenAI Codex, which all give the embedded browser more rights than a normal browser. The bugs led to code execution in Cursor, file theft in Claude Desktop, Codex and VS Code, and address-bar spoofing in Copilot. Cursor and Copilot are patched; OpenAI and Microsoft called the reports a workspace trust boundary and left Codex and VS Code open.
cyber
CrowdStrike links VAPOR PANDA to AI-agent attacks on over 200 organisations
CrowdStrike attributes the intrusions to a China-nexus actor that ran its ARTEX framework against the DeepSeek model and a second commercial model. In eleven days the campaign targeted more than 200 organisations in at least 14 countries. Nine of those were breached.
research
Anthropic: experts adopt an agent's coin-flip preference as their own judgment
Anthropic gave 50 experts four analytical tasks in which the agent received a random preference between two equal options. Participants followed that preference in up to 86 percent of cases and most treated the outcome as their own conclusion. Experience, expertise and trust in AI made no difference.
business
OpenAI doubles Pro usage limits in its competition with Anthropic
Limits for Work and Codex double on Pro 100 and Pro 500. On Pro 200 the lower bound doubled and a planned cut was reversed after customer complaints. Existing subscribers get the new quotas automatically and Plus is unchanged. OpenAI notes that double the tokens does not mean twice the finished tasks.
cloud
AWS revenue rises 39 percent to $46 billion while $220 billion capex drains cash flow
Amazon reported AWS revenue up 39 percent to $46 billion and put capital spending for the full year at $220 billion. The stock fell about 7 percent in after-hours trading.
markets
Global AI selloff pushes Taiwanese stocks toward a technical correction
Taiwan's benchmark index heads for a correction as the global selloff in AI shares deepens. Bloomberg links the move to weaker big-tech results and high capital spending in the sector. Asian chip stocks and Intel are under pressure as well.
policy
Trump declares robot imports a national security threat
The US president designated imported factory robots a national security threat and ordered trade talks, with tariffs as leverage. The decision extends the tariff pressure that already covers semiconductors and AI hardware.
platforms
Reddit tests login-only access outside the US as AI eats search traffic
Reddit starts tests in which only logged-in users outside the US can see content. CEO Steve Huffman cites the loss of search traffic to AI answers as the cause. He also hints at a more consistent home feed and a possible incognito mode.
open source
Prime Intellect makes DeepSeek's fused Mega MoE kernel trainable
Prime Intellect published a backward pass for DeepSeek's fused Mega MoE kernel, so the kernel now handles training as well as inference. Qwen3-MoE training steps in prime-rl speed up by up to 34 percent, with a 3.11x faster forward pass.
Key Trends
- • Agent credentials outrun agent oversight: a personal Claude agent sat inside a Pentagon network for two months.
- • The coding assistant is the attack surface: five tools share the same embedded-browser weakness.
- • Offensive agents scale: one actor reached 200-plus organisations in eleven days.
- • Judgment follows the model: experts adopt a random preference in 86 percent of cases.
- • AI capital spending has become a market risk, not only a balance-sheet item.
- • Efficiency work moves to open kernels: 3.11x on the forward pass lowers the cost per training step.
Wat vond je van dit bericht?
