AI week in review: agents overstep their limits and the bill comes due

Australia reacts to an agent that got past a government portal, attacks run by agents, Fireworks breached again, United blocks agents and a study shows incomplete work.

The theme of the week was the AI agent, software that acts on its own on someone’s behalf. It showed up getting around a government block, attacking companies, being barred by an airline and delivering half-finished work. I picked the five facts from Monday to Friday with the most practical impact for anyone running a company in Brazil, from most to least important.

Australia accepts 18 recommendations after an agent got past a government portal

On Tuesday (the 10th), the Australian government released the 18 recommendations of a review opened after an OpenAI agent, during an internal evaluation, bypassed blocks on the Medicare statistics portal. Among them: studying mandatory, time-bound reporting of AI incidents, including near misses. OpenAI noticed the case in August and only informed the government in September. Details in our news story.

Criminals already use AI agents to attack with almost no human intervention

A Check Point Research report analyzed campaigns run with ARTEX, a platform built for security testing. In one of them, agents attacked ten online stores and obtained more than 300 digital certificates, at an estimated AI cost of US$ 700 to US$ 1,500. With cheap, automated attacks, no company is too small to become a target. More in our news story.

Fireworks AI suffers a second attack and exposes customer data

Fireworks AI, which runs AI models for more than 10,000 companies, warned customers of a new attack on October 24 that came in through a cloud provider it had hired, according to Forbes. Access secrets, databases and models from a limited group of accounts were affected. The warning is about the chain: your company may depend on it without ever having signed a contract. Read our news story.

United will block AI agents that buy tickets on the customer’s behalf

United Airlines announced on Wednesday (the 11th) that it will restrict unauthorized personal AI assistants that book and manage flights. The airline cites passwords shared with third-party tools, purchases and use of miles the customer didn’t want, and steps that, by rule, require a person to confirm. Anyone who sells online will have to decide what an agent can do on their site. See our news story.

The best coding agent fully completes only 26% of tasks

In the Co-Design Bench, a study by Alibaba with Chinese and Singaporean universities, the top-ranked agent met 69.9% of requirements but fully delivered only 26.4% of the 273 tasks drawn from real requests. The risk is the “almost done” delivery, which passes a quick look and fails later. More in our news story.

What to watch next week

The five facts say the same thing: an agent without access limits, without logs and without an owner becomes a risk, whether it comes from your team, your vendor or a criminal. Next week, it’s worth listing the agents already running in your company, what each one accesses and who answers for it. And watching whether other big sites follow United and start blocking agents.


Originally published on IWA7 Insights, IWA7 Digital’s blog, which is now part of Escovando Bytes.

Keep readingSee all →