Researchers question Anthropic’s claim of a ‘90% autonomous’ AI cyberattack
Researchers are pushing back on Anthropic’s claim that a recent cyberattack was 90 percent automated by its Claude AI system.
Anthropic said a Chinese state-linked group used Claude Code to automate most steps of a hacking campaign against more than 30 targets. But outside experts say the results look far less dramatic. Only a small number of attacks succeeded, and researchers note that the tools used were standard open source ones that defenders already know how to spot.
Security experts also question the idea that attackers can get AI models to behave with near perfect compliance. They point out that developers and white-hat hackers still struggle with models that hallucinate, refuse tasks, or give inconsistent results. Even Anthropic admitted Claude often made mistakes, overstated findings, and fabricated credentials during the operation.
Anthropic says the attackers bypassed guardrails by splitting tasks into small steps and framing them as defensive exercises. The company’s report describes a multi-stage workflow where Claude handled scanning, credential checking, and movement across systems, with humans stepping in only at key points.
Researchers agree AI can speed up some hacking chores such as triage or log review. They just aren’t convinced this incident marks a major leap in attacker capabilities. For now, AI-assisted hacks still look more like incremental efficiency gains than fully autonomous cyberattacks.
Supported by Rocket
If you’ve used AI code tools, you know the frustration. Vague prompts cause unintended changes. Ask for a navbar update, get footer modifications you didn’t want. Spend time fixing what AI broke.
Rocket’s Precision Mode solves this with structured commands. Over 100 commands covering your entire workflow. “Connect Supabase database” handles auth, tables, and API keys. “Fix responsive breakpoints” auto-repairs layout issues. “Refactor component structure” reorganizes code cleanly.
No interpretation layer. No guessing. Commands execute exactly what you specify. Works with NextJS, React, and a Flutter.
First-try accuracy: 95%. Build time: cut by 10x. Zero prompt engineering required.
ElevenLabs teams up with celebrities for AI voice deals
ElevenLabs has signed Michael Caine and Matthew McConaughey to create authorized AI versions of their voices. It’s the latest sign that Hollywood’s relationship with AI is shifting after years of tension and the strike-era debates around digital likeness and guardrails.
Meta took an early step last year with AI assistants based on Kristen Bell and Judi Dench. Now McConaughey, who is also an investor in ElevenLabs, will use his AI voice to translate his newsletter into Spanish audio.
ElevenLabs is also launching a new marketplace where brands can license approved AI voices from well-known personalities. Alongside Caine and McConaughey, the lineup includes Liza Minnelli and Dr. Maya Angelou.
The company has quickly become one of the leading AI audio startups, backed by investors like a16z and ICONIQ, and is positioning itself as the go-to platform for celebrity-authorized AI voice work.
ChatGPT rolls out pilot group chats in four countries
OpenAI is testing a new group chat feature for ChatGPT in Japan, New Zealand, South Korea and Taiwan. The pilot lets users collaborate in shared conversations on both mobile and web, and is available to Free, Plus and Team users.
OpenAI says the goal is to learn how people actually use group chats inside the app before expanding more widely. Early testers will be able to send feedback to shape the final version.
Group chats are invite-only, support up to 20 people and keep personal ChatGPT memory and private chats fully separate. Users can leave at any time and the group creator can only leave voluntarily. For anyone under 18, content filters and parental controls remain on.
Starting a group is simple. Tap the people icon, add participants directly or send a link. Adding someone to an existing one creates a new group so older chats are not altered.
ChatGPT in groups works like a normal conversation. GPT 5.1 Auto handles responses with search, image generation, file uploads and dictation. Usage limits only apply when ChatGPT replies. You can tag it when you want input, and it can react with emojis or generate images using group members’ profile photos.
This is the latest sign OpenAI is moving ChatGPT toward a more social, shared platform. The company recently launched Sora 2, a TikTok-style app for AI video creation and sharing.
Anthropic releases open tool to measure political bias in chatbots
Anthropic has published an open source method for rating how politically evenhanded AI models are. The company says Claude Sonnet 4.5 and Claude Opus 4.1 scored about 95 percent on its test, which checks how well a chatbot presents opposing viewpoints and how often it refuses to answer. That puts Claude ahead of Meta’s Llama 4 and GPT 5, but slightly behind Google’s Gemini 2.5 Pro and Elon Musk’s Grok 4.
The tool arrives as pressure grows from regulators and politicians. President Trump’s new Woke AI executive order requires chatbots used by the government to stay politically neutral, but the definition of neutrality is vague and often contested. Studies already show chatbots can appear biased even when they are aiming for factual accuracy.
Anthropic says there is no agreed standard for measuring political bias and hopes releasing its method publicly will help the industry develop shared benchmarks.