Agents Got Cheaper. Attacks Got Cheaper Too.31 posts
← Back to InsightsAI Trends

Agents Got Cheaper. Attacks Got Cheaper Too.

September 13, 2026·6 min read

This week: OpenAI opened its Agents API in public beta on 10 September, putting the Codex harness — session management, tool orchestration, sandboxing, subagents — behind a single API call and one bill. Twenty-four hours later, Anthropic's fourth Threat Intelligence Report documented AI-orchestrated espionage against 20+ organisations by a Russia-linked group, biological research misuse, and a pattern the lab now calls "vibe hacking" — general goal in, script iterations out. Sakana AI released Fugu Ultra v2 on 11 September, a multi-agent router that beats frontier models on five of eight benchmarks at US$5/US$30 per million tokens. Closer to home, Meta's WhatsApp Business Platform confirmed 1 October pricing: service and utility replies inside the 24-hour service window become billable in Malaysian ringgit — the free-conversation era ends in 17 days.

Here's what happened, why it matters, and what your business should do about it.


The Big Three

1. OpenAI put the Codex harness behind one API call

On 10 September, OpenAI opened the Agents API in public beta. The pitch is narrow and honest: the same harness that runs Codex — model calls, tool routing, conversation history, context compaction, subagent coordination, a Linux sandbox for editing files and running code — is now a service you rent instead of code you maintain. There is no separate harness fee; you pay for model tokens, tools, and sandbox minutes.

For teams that have tried to build agents in-house, this collapses a category of engineering debt. Context compaction alone — deciding what to summarise, what to keep verbatim, and when to hand off to a subagent — has burned through six-figure ringgit budgets at companies who wrote it themselves and then watched it drift. OpenAI is now the one drifting it.

So what: if you were quoted RM80,000 to RM300,000 for a "custom AI agent" earlier this year, the underlying complexity of that quote just moved to a rented layer. That does not mean the price should fall by the same amount — your integrations, prompts and evaluation still cost real hours — but it does mean the harness plumbing stops being a defensible line item. If a vendor still charges you for that piece specifically, ask what they are building that OpenAI is not.

2. Anthropic's threat report caught up with the agentic era

Anthropic released its fourth Threat Intelligence Report on 11 September, covering December 2025 through August 2026. The headline finding: attackers now hand a general goal to a model and let it survey the target, write scripts, run them, summarise findings and iterate — a pattern the lab labels "vibe hacking." Documented cases include a Russia-linked cyber-espionage campaign automating intrusions against more than 20 organisations, systems designed to identify dissidents, and attempts to use the model for research involving viruses and toxins.

Two things worth separating. First, the report is Anthropic's own product data — one lab's window into misuse of its own model, not the whole picture. Second, the fixed-point observation matters regardless of source: the same "agent that takes a goal and gets on with it" capability that makes the Codex harness worth renting is also what makes an intrusion loop worth writing.

So what: if you run an SME with any customer data — a CRM, an accounting cloud, a webshop backend — the attack economics against you changed this year. It costs less to try you. It does not cost less to defend you. This is the moment to enable multi-factor authentication on every login your business uses, review who still has access after leaving, and confirm your Xero/QuickBooks/accounting tool has audit logging turned on. None of that is glamorous. All of it is what the report actually recommends between the lines.

3. Sakana AI shipped an orchestration model that beats frontier labs on price

Also on 11 September, Sakana AI released Fugu Ultra v2 — a model that routes tasks across a pool of open and specialised models and recursively calls itself. Sakana's own evaluation, which should be read with the usual scepticism reserved for vendor benchmarks, puts it best or joint-best on five of eight tests, including 48.3 on Chartography against Claude Opus 5 at 27.3. Pricing: US$5 per million input tokens, US$30 per million output, US$0.50 cached — rising to US$10/US$45/US$1 above 272K token context.

The pattern is worth noticing more than the model. GPT-6 Astra sits at US$10/US$50; Claude Fable 5.1 at the same US$10/US$50; Fugu Ultra v2 undercuts both by orchestrating cheaper models underneath. Cost is now something you can architect around, not just negotiate.

So what: if a vendor tells you their solution "uses the best AI" without saying which model or letting you swap it, you are locked into their markup. Insist that any AI-heavy quote name the models it uses and specify what happens if a cheaper one becomes adequate. Six months ago that question was pedantic. Today it separates operators from resellers.


Closer to Home: Malaysia

WhatsApp is about to bill for a conversation you have been having free. Meta confirmed this week that from 1 October 2026, service and utility replies sent inside the 24-hour customer service window on the WhatsApp Business Platform become chargeable. Utility messages will cost roughly RM0.0564 each; service messages become per-message billable with no volume discount. Malaysian ringgit billing has been available on the platform since April, so nobody can plead currency confusion on 1 October.

For an F&B business handling 200 order-status replies a day, that is roughly RM11 in daily platform charges before the Meta Business Agent tokens Meta priced separately at US$2 per million (about 4–5 sen per conversational exchange). For a trading SME running WhatsApp catalogues, quotations and confirmations, monthly cost exposure moves from "effectively free" to a real line item you should be modelling now, not in October.

Two practical steps this month. First, audit the last 30 days of your WhatsApp Business activity: how many replies land inside the 24-hour service window, how many are automated, and how many are pure human back-and-forth. Second, decide which of those are worth automating with a template message versus keeping as a human touch — because the pricing structure will start rewarding one over the other from day one.

Separately, the e-invoice compliance clock has not moved since PM Anwar's 1 September threshold change. Businesses under RM3 million in annual turnover are exempt from mandatory rollout, but the RM10,000-per-transaction rule is still live, the Phase 4 grace period still ends 31 December 2027, and LHDN's February disclosure of 500,000+ non-compliant cases and RM1.4 billion unreported income has not been superseded by any better number. Do not confuse exemption with amnesty.

Singapore note: no material SME-relevant policy movements this week.


What This Means for Your Business

  1. Model your WhatsApp bill before 1 October. Pull a month of your WhatsApp Business conversation logs, classify them by category (marketing, utility, authentication, service), and multiply by Meta's published rate. If your annual exposure is above RM3,000, that is a spend line worth designing around. If it is above RM15,000, it justifies re-plumbing your first-response flow with templates and structured buttons rather than free-text replies.
  2. Ask any AI vendor to name their models and let you swap them. Fugu Ultra v2 is the third launch in two months that makes cost architecture matter. A vendor who cannot tell you which model they route to, or who charges you a fixed price when the underlying token cost keeps falling, is charging you for opacity. Real operators are transparent about their model stack because they know it will change.
  3. Turn on the boring security controls this week. Multi-factor authentication on every business login, revoked access for anyone who left in the last 12 months, audit logging on your accounting and CRM. Anthropic's report is not a marketing document; it is a description of what attack loops look like now, and none of the mitigations require anything you cannot do yourself before the end of the month.

The Practical Question

If WhatsApp charged you for every reply starting tomorrow, which conversations in your business would you automate first — and which would you protect as human?

The answer separates SMEs that treat WhatsApp as a phone from SMEs that treat it as a channel. There is no wrong answer in the abstract; there is a wrong answer for your specific mix of order confirmations, delivery updates, quotation back-and-forth, and complaint handling. What matters is that you have thought about it before 1 October, not on 15 October when the first invoice arrives.


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.

Talk to us →