Claude Joined Slack. Washington Pulled the Plug.
Three different doors opened this week — into your team chat, into your marketing budget, and onto a classified server.
“On 23 June, Anthropic launched Claude Tag, a public beta that drops Claude into Slack as a tagged, shared teammate for Enterprise and Team customers — and disclosed that 65% of its own product team's code is now written by an internal version of the tool. The same day, the Associated Press reported that NSA director General Joshua Rudd told a Senate panel Anthropic's Mythos model "broke into almost all of our classified systems, not in weeks, but in hours," after which Washington restricted Mythos 5 and Fable 5 to U.S. citizens. And at Cannes Lions, OpenAI made its first appearance as an advertising company — disclosing that its ads pilot has crossed US$100 million in annualised revenue in under six weeks, against 900 million weekly users of which roughly 20% ask questions with direct commercial intent.”
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. Claude Took a Permanent Seat in Slack
Anthropic's Claude Tag, in public beta from 23 June for Claude Enterprise and Team plans, isn't a chatbot in a browser tab anymore — it's a tagged participant inside the channels your team already lives in. Mention @Claude and it picks up the thread context, breaks the request into stages, and reports back. Anthropic's own product team says 65% of its internal code is now written by their version of the same tool, which is a credibility marker, not a feature.
For an SME owner, the boring part is the important part. The distance from "AI tool we should pilot one day" to "colleague we already added to the operations channel" just collapsed. If your team doesn't use Slack, the principle is unchanged: AI vendors have stopped asking you to switch tools and started moving into the tools you're already in.
2. Anthropic's Top Model Ran Through Classified Systems in Hours — and Got Pulled
Reported by the Associated Press on 23 June and confirmed at the Senate Banking Committee hearing on 11 June, General Joshua Rudd, head of the NSA and U.S. Cyber Command, said Anthropic's Mythos model "broke into almost all of our classified systems, not in weeks, but in hours" during a closed test under Anthropic's Project Glasswing programme. The U.S. administration responded by ordering Anthropic to block all foreign nationals from Mythos 5 and Fable 5. Anthropic disabled both models for every customer to comply, with paying Fable 5 subscribers effectively losing access mid-billing-cycle. Over 100 cybersecurity leaders, including signatories from Adobe and Nvidia, wrote back that the models are "quite good" at finding exploits but "not uniquely good."
For a Malaysian business that bought a Claude subscription this year, the immediate read is unsentimental: governments will now switch off frontier-tier AI for non-citizens with no notice and no refund. If you've built an automation on a single top-end model from a single vendor, you've also built a single point of failure with a foreign regulator's hand on the switch. The fix is portability, not panic — design your AI workflows so the underlying model can be swapped for a cheaper or different one in an afternoon. The good news for most SME use cases is that you almost certainly weren't on the frontier tier to begin with.
3. OpenAI Came Out as an Advertising Company
OpenAI made its first appearance at Cannes Lions this month and used it to confirm what its product moves had been signalling: advertising is now a core revenue line. The company disclosed that its ads pilot has reached more than US$100 million in annualised revenue in under six weeks, against 900 million weekly active users, and said internal projections target roughly US$2.5 billion in 2026 ad revenue scaling toward US$100 billion by 2030. Around 20% of ChatGPT queries, the company said, carry direct commercial intent — "where should I eat tonight," "what laptop should I buy."
For any business that depends on being found online, the implication is direct. For two decades the discovery question was "do we rank on Google?" The second answer — "are we visible when 900 million people ask a model where to buy what we sell?" — is now a live one, with billable inventory behind it. For a Malaysian SME on Shopee, Lazada or its own site, the practical move is to make sure product names, descriptions and FAQs read cleanly to a model: clear category words, plain pricing, location, what you do. Write like a human being for a machine that reads like one.
Closer to Home: Malaysia
No Putrajaya headline this week — so the most useful thing this column can do is read the clock honestly.
Eighteen months of Phase 4 grace period remain (to 31 December 2027) for businesses with RM1–5 million revenue, with penalties of RM200–RM20,000 per invoice starting 1 January 2028. Subtract the realistic 3–6 months a clean integration takes and the unhurried start window closes around the end of this year. The RM10,000 individual e-invoice rule is already live and already enforced — no consolidation allowed on any single transaction at or above that figure, regardless of which phase your business sits in. LHDN's February disclosure of 500,000+ non-compliant cases and RM1.4 billion in unreported income has not been retracted; the grace period suspends penalties, not audits.
The MSME Digital Grant MADANI continues to offer 50% matching funds up to RM5,000 for tools bought from MDEC-listed providers — and the SJPP RM5 billion financing guarantee (80% coverage, tenures to 10 years) that came with the April extension hasn't moved. If the digitalisation project was waiting on cash, the funding stack is sitting where Putrajaya left it.
Singapore note: OpenAI's Singapore Applied AI Lab continues to staff up its Forward-Deployed Engineer roles — agentic-deployment expertise concentrating in the city-state will leak across the Causeway long before it shows up on a Malaysian procurement list.
What This Means for Your Business
1. Add AI to one channel your team already uses
Don't run a vendor pilot. Pick the chat, ticketing or document tool your staff actually open every morning and turn on its AI features for one week. Slack, Microsoft Teams, Google Workspace, your CRM, your accounting package — pick one. Measure the time saved on one named task. That's the entire experiment.
2. Rewrite your top 20 product pages for the model, not the search bar
Pull the 20 listings, services or location pages that earn the most enquiries. Read each one aloud and ask whether a model could answer a buyer's question about your product from that text alone — price, category, availability, who it's for, what makes it different. If it can't, rewrite it in plain sentences. The same edits help your existing SEO; they also put you in the answer when the question is asked of a model.
3. Map your single points of failure on AI
List every business workflow that now depends on AI — drafting, classifying, summarising, lookup. Note which vendor and which model each one calls. Anywhere a workflow can only run on one specific top-tier model, write down a fallback. This week's news made the cost of skipping that exercise concrete.
The Practical Question
If a buyer in Penang asks an AI tonight where to find what we sell — does our website answer the question, and does our system know what to do when the order arrives?
The discovery layer and the operations layer are both being rebuilt at the same time. Your competitors who get found by the model and then drop the ball at the e-invoice will lose the customer twice. The owners who win the next eighteen months will be the ones whose listings read well to a machine — and whose back office can actually deliver, compliantly, on what the machine sent them.
At The Empyrean, we help Malaysian SMEs find the practical, repeatable tasks where AI delivers value without disruption. If you're not sure where to start, we're happy to take a look at your operations and tell you honestly what would make sense.