AI gets more keys to your digital life, and attackers are keeping pace

A theme runs through this week's news: the more AI is allowed to do on our behalf, the more it matters who holds the keys.

When the helper becomes an entry point

The security firm Forever Security is warning about an attack in which a website drops a hidden instruction file into the Downloads folder, which AI coding agents then follow. The firm itself reports the success rates from its own tests. Another example is the popular tool LiteLLM, where version 1.107.0 was compromised with malicious code on October 21 that steals credentials and gives the attacker remote access.

Attackers are also using AI themselves. The security firm Unit 42 describes a low-skilled criminal who used AI as an adviser to attack two Indian universities, and who, according to the firm, could hardly have pulled it off without it. Straiker says the CyberCartel group used several AI models against more than 400 banks in Latin America, but that is the firm's own investigation.

Defenses are being built out

Fortunately, tools are also arriving from the other direction. 1Password has launched a solution in which AI agents can log in to your accounts without seeing the actual password, and you approve each request yourself. Microsoft plans to deploy so-called Guardian Agents, meaning AI guards, to monitor other agents, because agents often have more permissions than necessary. Anthropic's Claude Security, which is in public beta, now continuously scans code and marks bugs as fixed.

Privacy is the other front. OpenAI lets you save selected ChatGPT conversations behind a four-digit PIN. Meta is working on a private mode in its Muse AI, where conversations are supposed to be hidden from Meta itself, but critics say Meta has not shown how it is supposed to work. Meanwhile, Claroc says that more than 6,800 addresses to shared Muse conversations can be found by strangers, some containing personal information. According to one analysis, Google's new privacy rules from October 16, 2026 could mean that data from Gmail, Calendar and Photos affects personalization if you connect them to Google Search.

Authorities and mixed signals

AI companies are also under political pressure. Texas has sued OpenAI and Sam Altman for allegedly misleading the public about ChatGPT's safety, and is demanding, among other things, that ChatGPT not be offered to minors in Texas. At the other end of the spectrum, the UK is betting on AI as a weapon against biological threats, with a new laboratory and GBP 100 million in government funding. In Derbyshire, the watchdog IOPC is investigating whether a detective used Microsoft Copilot inappropriately in criminal cases, but these are only allegations.

The picture is not one-sidedly negative. A Nigerian study of 371 adults found no significant differences between ChatGPT and clinicians on cancer information. On the other hand, a study shows that chatbots forced to pick a candidate choose a side, even though the consumer versions refused to answer.

What can you do?

Don't give AI tools more permissions than the task requires. Check whether conversations you share can be found by others. Review your privacy settings when the rules change. And treat files and packages from the web with the same caution, whether it is you or your AI assistant who opens them.