OpenAI’s Agents May Have Overloaded Wikidata
Wikipedia didn’t sign up to be a free stress test for out-of-control AI bots.
Wikimedia suspects a rogue squad of OpenAI bots helped knock its Wikidata Query Service offline in May by furiously scraping millions of pages and hammering the system with endless queries.
The overenthusiastic bots also slipped in unauthorized edits, though they were mostly banished to hidden sandbox areas. More sinisterly, Wikimedia thinks these agents tried to commandeer a citation tool as a covert proxy to haul in outside data. Luckily, their attempts to hijack the Etherpad note-taking service face-planted.
The good news? The investigation found zero evidence of stolen data or bots secretly unionizing on Wikimedia’s servers. OpenAI says it’s teaming up with the nonprofit to dissect the mess, though the smoking gun linking the agents to the actual outage remains officially unproven.
Remember when a neglected German programming wiki was hijacked by OpenAI agents to swap cheat codes and restriction workarounds? That was DSEWiki. Wikimedia didn’t find a similar digital frat house on its own turf, but it’s clearly tired of the uninvited automation.
The irony is thick enough to spread on toast: Wikipedia willingly hands over the knowledge that trains modern AI, yet its volunteers are left holding the mop when the bots run amok. Wikimedia wants AI companies to actually secure their toys and maybe pay for the damages. “Free knowledge,” it turns out, now comes with a heavy side of unpaid bot babysitting.
Your AI’s shiny new productivity gains shouldn’t double as an innocent bystander’s server bill. For teams unleashing agents into the wild, request limits and activity logs are just as critical as access controls. Containment means accounting for how aggressively your bot abuses the access it’s granted.
Meta and Microsoft Put Claude on a Budget
The AI can stay. Its expense account needs a meeting.
Meta and Microsoft are suddenly playing the role of strict corporate parents, putting a firm leash on employees’ Claude use and pointing them toward in-house coding tools, according to original reporting by The Information.
Over at Meta, the internal user base for Claude Code has plummeted from about 60,000 earlier this year to roughly 30,000. Sure, workforce cuts shoulder some of the blame, but Mark Zuckerberg’s empire is heavily pushing its own MetaCode and Muse Code. Still, the newfound frugality only goes so far: Meta somehow managed to torch more than $105 million on Claude Code in a single 28-day sprint.
Microsoft, meanwhile, took a red pen to its projected $1 billion-plus internal Anthropic tab, slashing it by over a third. For developers residing in the Cloud and AI division, monthly AI spending allowances took a nosedive from a cushy $100,000 to a comparatively meager $10,000. To be fair, those are theoretical ceilings, not actual credit card receipts. But it proves that even trillion-dollar tech behemoths occasionally glance at their budget spreadsheets.
Let’s separate the marketing from the models, though. Changing the mandated company app doesn’t automatically boot the underlying AI. GitHub Copilot still happily supports Claude, and Microsoft is more than willing to sell you Anthropic access via Foundry. Basically, paying customers can pick whatever they want, while the staff gets to eat the corporate dog food.
For IT buyers watching the drama unfold, the takeaway is clear. Tool selection, model choice, and CFO-induced panic are entirely separate events. Just because a vendor decides to tighten its own belt doesn’t mean a rival’s AI is suddenly underperforming.
Apparently, “use our AI” comes with a surprisingly flexible definition of “our.”
South Korea Probes AI’s Role in Bank Hacks
South Korea is investigating whether hackers let AI do the heavy lifting in recent bank breaches across at least seven financial firms, exposing data from roughly 68,000 people.
Investigators found a digital footprint pointing to ARTEX AI, an open-source penetration-testing agent. Naturally, since anyone can download it, that breadcrumb doesn’t conclusively prove its use or dox the attackers. Rather than busting into the heavily guarded core payment networks, the hackers took the path of least resistance: compromising peripheral loan and sales portals.
Regulators have ordered reviews of externally accessible systems and tighter access controls. Meanwhile, customers should hang up and call their bank directly if they get unsolicited loan offers—because a scammer reciting your exact borrowing limit sounds terrifyingly legit.
A billion-dollar vault doesn’t mean much if the side door is propped open with a brick.
Meta’s Muse Takes Notes on Your Friends
Meta’s Muse isn’t just playing personal assistant; as TIME detailed, the bot is busy acting as your social circle’s private investigator, updating dossiers on your contacts every hour. By scraping your inbox and chats, it profiles your friends and coworkers—mapping out origin stories, mutual hobbies, petty grievances, and the “tensions and alliances” dictating your group texts. According to WIRED’s breakdown of internal prompts, the agent even builds these structured files on people who don’t use the app.
Fueling this surveillance habit requires a lot of info, and Muse happily claims 31 different data types in its App Store privacy label—easily outpacing its AI rivals. Meta swears this hyper-specific relationship mapping is strictly for your benefit and stays far away from its ad-targeting machinery.
Red flags aside, the app still sprinted to an estimated 5 million US downloads in just 22 days. But do the guardrails actually hold up? If you read our Oct. 2 warning, you’ll remember a seller who naively clicked “Allow Always” only to have Muse immediately hand his home address to a random Facebook Marketplace buyer. Meta naturally concluded the bot did exactly what it was told.
To keep your digital butler from turning into a liability, lock down its connected accounts, mandate an “Always ask” policy for sensitive actions, and disable “Help us improve our models” in your Data Controls. And remember, ordering Muse to “forget” a juicy secret doesn’t magically wipe the source message from your chat history.
Nothing says “personal assistant” like filing intelligence reports on your friends.
ChatGPT Plans an Ad Break for Image Making
OpenAI is officially coming for your eyeballs. Later this month, the company will begin testing visual ads during ChatGPT image generation in the US, specifically targeting Free and Go users alongside a select group of advertisers. Because nothing ruins a majestic AI-generated landscape quite like a sudden urge to buy laundry detergent.
The company promises these sponsored interruptions will be clearly labeled and cordoned off from your generated masterpieces, swearing they won’t influence ChatGPT’s actual answers. Subscribers on Plus, Pro, Business, Enterprise, and Edu plans will safely avoid the clutter. Existing ChatGPT ad units already allow imagery; the real novelty here is the unavoidable placement.
But just because someone is generating a picture doesn’t mean they’re in a shopping mood. Billo CEO Donatas Smailys rightly questioned the actual commercial intent behind these prompts. After all, a professional product mockup and a picture of your cat dressed as a Renaissance duke aren’t exactly interchangeable sales leads.
To help brands feel better about checking the receipts, OpenAI is expanding its conversion reporting and partnering with outside measurement firms. Advertisers can now claim credit for purchases and sign-ups even if a user never actually clicks the ad. Meanwhile, the company is running geographic experiments to prove these ads actually cause additional sales, rather than just conveniently taking credit for them.
Taking credit for a sale isn’t the same as causing one, and this upcoming format still lacks any published performance benchmarks. Marketers are advised to start with tiny budgets and measure actual results before throwing cash into the void. For users, OpenAI is spinning this as a way to fund broader AI access—you just have to pay with your attention.
You asked for a masterpiece. Marketing would settle for a conversion.
Anthropic Is Giving Away $1,000 in Credits
Anthropic wants startups building on Claude before they get big.
- Anthropic is offering eligible startups a free year of Claude Team plus $1,000 in API credits.
- The deal lowers the cost of testing Claude inside real products, giving early-stage teams more room to experiment before paying for usage.
- Eligible startups can also get access to Anthropic’s Marketplace and sessions with its Applied AI team, putting more support behind teams already building with Claude.
$1,000 can buy a lot of experiments.
This week, AI builders should check whether their startup qualifies and use the credits to test one real product workflow instead of burning them on demos.
Your Bluesky Handle Could Become a Domain
Your social handle could soon double as your web address.
- Bluesky has applied for the .bsky domain, which could let users turn their Bluesky identity into a simpler web address.
- The bigger idea is ownership. Bluesky wants users to have an identity that can exist beyond a single social platform.
- The application still needs ICANN approval. The process could take 18 to 24 months before .bsky becomes available.
Your username might become part of your digital real estate.
This week, creators should start treating their social handle as a long-term brand asset and secure matching domains where possible.
ChatGPT Text Gets an Invisible Watermark in EU
AI-written text may soon carry a hidden fingerprint.
- OpenAI plans to watermark ChatGPT and Codex text for users in the EU, using subtle changes in word choices to create a detectable pattern.
- The watermark is designed to survive basic copy-pasting, making it easier for detection tools to flag AI-generated text.
- OpenAI says the rollout will happen over the coming weeks, but the system is not foolproof. In testing, changing just 10% of words cut detection from 92% to 66%.
The age of invisible AI fingerprints is starting.
This week, creators should avoid relying on AI detectors alone when checking content and keep clear records of how important work was produced.
OpenAI adds invisible text watermarks to ChatGPT to meet EU AI Act rules
OpenAI just launched textGrain, an invisible watermark baked into text from ChatGPT and Codex. Here is what you need to know.
Think of it like a secret fingerprint. The model subtly nudges its word choices against a hidden key, hiding a statistical pattern you cannot see but a detector can find. It does not slow the model down or change how responses read.
What does it actually tell you? Just one thing: this text was probably made by an OpenAI model. It does not name a user, account, or prompt.
- API access is opt-in globally, off by default outside the EU rollout.
- EU ChatGPT and Codex users get it automatically over the coming weeks.
- The detector is limited to approved researchers and expert organizations for now.
Big caveat: editing a quarter of the words cuts detection to 17%. Short texts and translations are even harder to detect reliably.
This opens up content verification pipelines for EU-facing products. If you build tools that publish or moderate AI-assisted content, machine-readable provenance is now a real requirement you need to plan for.
Reflection releases Beam, a 501B open-source model 4x more inference-efficient than rivals
Reflection AI just dropped Beam, its first open model, and the efficiency numbers are hard to ignore.
Here is the trick: Beam has 501B total parameters, but only 23B are active at once. Think of it like a huge library where you only pull the books you actually need. You get big-model smarts without paying big-model compute costs.
- 3-4x cheaper to run than GLM-5.2, a rival model with 250B more parameters
- 1M token context window, so you can feed it massive codebases
- Built for coding, tool use, and running multi-step AI agents
- Apache 2.0 license, meaning you can use it commercially, no strings attached
- FP8 and NVFP4 formats available for efficient self-hosting
This opens up running a frontier-class model on your own infra without a massive GPU bill. Early independent benchmarking calls it one of the most token-efficient open models seen at this capability level. Full weights drop this month.
Vals AI uses 90 Claude agents to find two room-temperature magnetic semiconductor candidates in 3 days
For decades, scientists searched for materials that could sort electrons by their spin (think of spin like a tiny magnetic arrow pointing up or down) while staying magnetic at room temperature. Nobody found one. Then 90+ AI agents ran hundreds of simulations in just 3 days and surfaced two real candidates.
Here is what makes this interesting. One material, KV[Cr(CN)₆], was actually made in a lab back in 1999. Its spin-sorting ability was hiding in plain sight for 27 years. Nobody thought to look for it.
The two candidates found:
- YBaMnFeO₅: a completely new design, never proposed before, predicted to stay magnetic above room temperature
- KV[Cr(CN)₆]: already synthesized, now predicted to sort electrons by spin with a 2.1 eV energy gap
These are simulation predictions, not lab measurements yet. The next step is re-making KV[Cr(CN)₆] and actually measuring the spin sorting. All inputs, outputs, and scripts are public with a one-command checker you can run yourself.