DeepSeek’s Super-Agent, ChatGPT account takeover, and the electric bill you didn’t expect
Compact Conversations for 2026-11-01: 7 AI stories, ai news worth knowing in just 5 minutes.
[Audio embed placeholder]
The Lead: DeepSeek open sources Super-Agent, a proactive coding agent
DeepSeek has released a research prototype called Super-Agent, a coding agent that proactively monitors developer group chats for coding issues and suggests solutions, positioning it against Anthropic’s Claude Code.
Why it matters: DeepSeek is explicit that this is a research prototype, not a finished product, and it’s open source, so teams can pull it down and see how it works.
Source: Bloomberg
Number to Know: Data centers consumed 4.4% of U.S. electricity in 2023, projected to reach 6.7-12% by 2028
Data centers consumed about 4.4% of all U.S. electricity in 2023, up from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023, and the Department of Energy’s Lawrence Berkeley National Laboratory projects that share could reach between 6.7% and 12% by 2028.
Why it matters: When a big data center connects to the grid, utilities may need new power plants and transmission lines, and depending on local rules, some of that cost can be spread across everyone’s electric bill, even if you never use AI.
Source: Tom’s Guide
The Feed
Parameter Labs describes ChatGPT account takeover via sandboxed iframes
Parameter Labs chained an exploit through ChatGPT’s sandboxed HTML previews to take over accounts, using a JavaScript library’s virtual keyboard and a YouTube endpoint to execute code on the page.
Why it matters: A crafted share link could exfiltrate a signed-in user’s private data or act on their behalf; OpenAI closed the hole and paid a $5,000 bounty.
Source: Simon Willison’s Weblog
Anthropic expands text watermark to four more Claude models, with a loophole on two older ones
Anthropic is watermarking Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 to meet the EU AI Act deadline, but on Opus 4.6 and Sonnet 4.6 the watermark only applies at default sampling settings.
Why it matters: A developer tweaking a single parameter can bypass the watermark on those two older models; Anthropic’s response is a compliance warning to keep settings at defaults.
Source: The Decoder
OpenAI plans post-midterms policy blitz with ‘speed limits’ and ‘AI compute campuses’
OpenAI will push for mandatory safeguards on powerful AI systems and rebrand its data centers as ‘AI compute campuses’ to shape the AI debate after the midterms.
Why it matters: Chief global affairs officer Chris Lehane told Axios the goal is that people always remain in control of the technology.
Source: Axios
Deepgram acquires Daily, maker of open-source voice AI framework Pipecat
Deepgram is buying Daily, the company behind the open-source voice AI framework Pipecat, for an undisclosed sum, and both companies say Pipecat stays open source and vendor neutral.
Why it matters: The acquisition gives Deepgram deeper real-time voice and video infrastructure expertise as the voice AI sector consolidates.
Source: TechCrunch
Interactive essay ‘Let’s Build a Brain’ explains neural circuits with simple simulations
Aatish Bhatia and Arkarup Banerjee walk through building a tiny creature called a spark with a simple brain, showing how different neural wiring produces different behaviors.
Why it matters: The essay demonstrates how apparent intelligence emerges from neural circuits, using just four neurons and two connections.
Source: Aatishb.com
One Thing to Try
Memdoor published a workflow where five small Qwen 27 billion parameter models tackle one programming problem as a team: a planner, two ideators, a solver, and a repair worker, each starting from a fresh context and talking only through files on disk. On a sample of hard LiveCodeBench problems, that five-model team matched or beat a single call to a much larger frontier model, at a median of about 40 minutes and 39 cents per run. The trick is the fresh-context discipline: every agent reads the files rather than relying on chat history, which keeps errors from compounding. It’s eight YAML files and no code, and you can pin any model you like, including a local one.
Sources
- DeepSeek open sources Super-Agent, a coding agent research prototype it says proactively monitors developer group chats for coding issues - Bloomberg
- You could be paying for AI through your electric bill — even if you never use it - Latest from Tom’s Guide in Ai
- How We Took Over ChatGPT Accounts - Simon Willison’s Weblog
- Anthropic’s watermark rollout leaves a loophole on two older Claude models - The Decoder
- OpenAI plans a post-midterms policy and PR blitz, pushing mandatory “speed limits” for powerful AI and rebranding its data centers as “AI - axios.com
- Enterprise voice AI tools provider Deepgram acquires Daily, which has raised $62M+ and develops open-source voice AI framework Pipecat, for - TechCrunch
- Let’s Build a Brain - aatishb.com
- Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard - GitHub
Transcript
Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.
Host A: [conversational] It’s Sunday, so this is our weekend update covering the last couple days, and today’s lead is DeepSeek. Bloomberg reports the Chinese AI lab has open sourced a coding agent research prototype called Super-Agent. The pitch is unusual: the company says it proactively monitors developer group chats for coding issues and suggests solutions, rather than waiting for a prompt.
Host B: [curious] That’s a different posture from most coding agents, which sit idle until you ask. DeepSeek is positioning Super-Agent against Anthropic’s Claude Code, the popular choice for terminal-based agents. Bloomberg’s Debby Wu reported the release, and DeepSeek is explicit that this is a research prototype, not a finished product.
Host A: [thoughtful] Right, and that prototype label matters. It’s open source, so teams can pull it down and see how it works, but nobody has independently verified its performance on real codebases yet.
Host B: [with emphasis] One number to know today: 4.4 percent. That’s the share of all U.S. electricity that data centers consumed as of 2023, according to figures Tom’s Guide pulled from the Department of Energy’s Lawrence Berkeley National Laboratory. The lab projects that share could reach between 6.7 and 12 percent by 2028.
Host A: [thoughtful] And here’s the part that lands even if you never touch AI. When a big data center connects to the grid, utilities may need new power plants and transmission lines, and depending on local rules, some of that cost can be spread across everyone’s electric bill. Tom’s Guide points to PJM, the grid operator serving 67 million people, where capacity auction prices rose roughly a thousand percent in two years, and Maryland households are estimated to be paying an extra 168 to 216 dollars a year.
Host A: [conversational] From the feed now, starting with security. Parameter Labs published a writeup, picked up by Simon Willison, describing how they took over ChatGPT accounts through the interactive HTML apps ChatGPT can render in a sandboxed iframe.
Host B: [curious] The attack chain was intricate. A JavaScript library OpenAI uses for graphs includes its own virtual keyboard, which could be triggered from inside the sandbox. The receiving code used innerHTML, letting untrusted code inject a new iframe without sandbox restrictions. From there, they abused a YouTube endpoint on OpenAI’s script allowlist to run JavaScript on the page itself, so a crafted share link could exfiltrate a signed-in user’s private data or act on their behalf. OpenAI closed the hole and paid a five thousand dollar bounty. [with emphasis] Willison’s own take is that the payout was a bit cheap for a vulnerability this serious.
Host A: [lighter] Next up, Anthropic. The Decoder reports the company is expanding its text watermark to four more Claude models to meet the EU AI Act’s December deadline. The watermark is an imperceptible pattern in word choice that Anthropic says doesn’t change meaning or quality, and it can’t be turned off. But on two older models, Opus 4.6 and Sonnet 4.6, it only applies at default sampling settings, so a developer tweaking a single parameter can bypass it. Anthropic’s response is a compliance warning to keep those settings at defaults.
Host B: In other news, Axios looks at OpenAI’s plans for after Tuesday’s midterms: a policy and PR blitz including proposals for mandatory speed limits on powerful AI systems, and a rebranding effort that would call its data centers AI compute campuses. Chief global affairs officer Chris Lehane told Axios the goal is that people always remain in control of the technology.
Host A: [quick aside] And TechCrunch reports Deepgram is acquiring Daily, the company behind the open source voice AI framework Pipecat, for an undisclosed sum. Both companies say Pipecat stays open source and vendor neutral.
Host B: [conversational] One Thing to Try: multi-model team solving with small local models. A project called Memdoor published a workflow where five small Qwen 27 billion parameter models tackle one programming problem as a team: a planner, two ideators, a solver, and a repair worker, each starting from a fresh context and talking only through files on disk.
Host A: [curious] On a sample of hard LiveCodeBench problems, that five-model team matched or beat a single call to a much larger frontier model, at a median of about 40 minutes and 39 cents per run. The trick seems to be the fresh-context discipline: every agent reads the files rather than relying on chat history, which keeps errors from compounding. It’s eight YAML files and no code, and you can pin any model you like, including a local one.
Host A: That’s Compact Conversations for Sunday. More AI news tomorrow. Until then, happy prompting.