How safe are wifi enabled talking toys?

Be very, very careful. It's not KRACK that is the problem, it is a lax attitude to security and privacy in general. So called "smart" consumer products can often be hijacked, accessed from the internet, or monitored. As a customer, it is hard to know if any specific product is safe or not.

The Norwegian Consumer Council has been on the case for a while, and produced a few horror stories. From a report, aptly titled #ToyFail, on three "smart" dolls:

When scrutinizing the terms of use and privacy policies of the connected toys, the NCC found a general disconcerting lack of regard to basic consumer and privacy rights. [...]

Furthermore, the terms are generally vague about data retention, and reserve the right to terminate the service at any time without sufficient reason. Additionally, two of the toys transfer personal information to a commercial third party, who reserves the right to use this information for practically any purpose, unrelated to the functionality of toys themselves.

[I]t was discovered that two of the toys have practically no embedded security. This means that anyone may gain access to the microphone and speakers within the toys, without requiring physical access to the products. This is a serious security flaw, which should never have been present in the toys in the first place.

And from an other of their reports, again aptly named #WatchOut, on "smart" watches for kids:

[T]wo of the devices have flaws which could allow a potential attacker to take control of the apps, thus gaining access to children’s real-time and historical location and personal details, as well as even enabling them to contact the children directly, all without the parents’ knowledge.

Additionally, several of the devices transmit personal data to servers located in North America and East Asia, in some cases without any encryption in place. One of the watches also functions as a listening device, allowing the parent or a stranger with some technical knowledge to audio monitor the surroundings of the child without any clear indication on the physical watch that this is taking place.

And the FBI agrees:

Smart toys and entertainment devices for children are increasingly incorporating technologies that learn and tailor their behaviours based on user interactions. These features could put the privacy and safety of children at risk due to the large amount of personal information that may be unwittingly disclosed.

So unless you have a real need (other than "this is cool") for these kinds of products, I would say that your best approach is to simply stay away from them.

It really depends on your threat model. I wouldn't be particularly worried about a particular sexual predator in your local area having the technical skills necessary to utilize Krack to inject voice into the toy. Unless it uses the vulnerable Linux driver, the key clearing won't work and the partial nature of the compromise for a general reset would make voice injection nearly impossible.

Similarly, as a client device, it doesn't offer a whole lot of security risk other than possibly as a listening device, depending on if it is always on or activated by pushing a button. Krack wouldn't make it usable as an entry point in to your network directly, so I don't see it as a particularly riskier device than any other IOT device.

As always in security, it comes down to your risk aversion though. Personally, if I thought it would be valuable to my child (who is also 3) I don't think I would consider the local security implications as a reason not to get it for my home environment. I'd be more concerned about the controls and security on the web side.

My main concern for IOT devices isn't the local compromise so much as the web connected remote compromise. The chances of a sufficiently skilled and motivated malicious individual in your direct proximity is pretty low. The chances of a motivated and malicious user on the Internet trying to remotely access the IOT device is significantly higher and it's important to understand what holes the devices punch in your network protections.

Also, as Michael was kind enough to point out, the interests of such a broad hacker are much less likely to be concerned with your privacy and much more likely to either be interested in attacks on your other computers or on the computational capabilities of the device as an attack bot.

Welcome to the Internet of Things(IoT). This is a... thing. Therefore, it can be assimilated

Mirai is a type of malware that automatically finds Internet of Things devices to infect and conscripts them into a botnet—a group of computing devices that can be centrally controlled.

And

One reason Mirai is so difficult to contain is that it lurks on devices, and generally doesn't noticeably affect their performance. There's no reason the average user would ever think that their webcam—or more likely, a small business's—is potentially part of an active botnet. And even if it were, there's not much they could do about it, having no direct way to interface with the infected product.

The problem is that security is seldom a consideration when making toys like this. The technology to make all this work is fairly simple, but the companies aren't paid to think about this. It's a child's toy. It's meant to be cheap and easy. And you get what you pay for.

Earlier this year, it was found that a similar child's toy had no security at all (emphasis mine)

A maker of Internet-connected stuffed animal toys has exposed more than 2 million voice recordings of children and parents, as well as e-mail addresses and password data for more than 800,000 accounts.

The account data was left in a publicly available database that wasn't protected by a password or placed behind a firewall, according to a blog post published Monday by Troy Hunt, maintainter of the Have I Been Pwned?, breach-notification website. He said searches using the Shodan computer search engine and other evidence indicated that, since December 25 and January 8, the customer data was accessed multiple times by multiple parties, including criminals who ultimately held the data for ransom. The recordings were available on an Amazon-hosted service that required no authorization to access.

I'm going to be honest. These things are scary powerful in what they can do. Even if it doesn't expose your messaging, it could still be used for something malicious like a DDOS attack. If I were you, I'd pass on anything like this unless there's something explicit about security.