Businesses are reporting benefits from artificial intelligence even as their data preparation and oversight processes remain incomplete, according to a survey by data and AI services provider Further.
An AI agent can finish an assignment and leave behind a second, less welcome deliverable: a question about who pays for what it did along the way.
- By John K. Waters
- 10/19/2026
OpenAI has disclosed four additional cases in which internal AI models bypassed restrictions or accessed systems without authorization while completing routine tasks, highlighting the difficulty of detecting misconduct as agents gain greater independence.
- By John K. Waters
- 10/19/2026
When an enterprise system develops a problem, the time spent fixing it can include waiting for someone to investigate, locating the right specialist, and securing approval to act. AI agents promise to shorten that process. For IT leaders, the question is how much work they can safely delegate.
Character.ai’s bankruptcy and proposed asset sale have left its future uncertain, prompting 40 attorneys general to seek safeguards for users’ sensitive conversations under any new owner.
- By John K. Waters
- 10/16/2026
Nearly half of the security and technology professionals surveyed by Salt Security confirmed or suspected that an AI agent had taken an unintended, unexpected, or unauthorized action in their organization over the past year.
Google is initially restricting Gemini 4 Argon to trusted cybersecurity partners before expanding availability.
- By John K. Waters
- 10/16/2026
Cybersecurity firm Sophos says unauthorized AI tools and agents are creating security visibility challenges for organizations.
Anthropic has suspended live internet access across its internal AI evaluations after finding that Claude models sometimes worked around restrictions to complete tasks, including exploiting software flaws and submitting forms on real websites.
- By John K. Waters
- 10/12/2026
OpenAI shelved GPT-6.1 Astra's planned October release after internal testing raised concerns about authorization and accurate reporting of its actions. Separate incidents and simulated evaluations show why recognizing a task boundary does not guarantee an AI agent will respect it.
- By John K. Waters
- 10/08/2026
Cybercriminals don't necessarily need to build their own malicious AI models. Increasingly, they're sharing prompts, jailbreak techniques, and access to existing models that can be incorporated into fraud and phishing operations.
Google’s newest AI model can help find the software flaws an attacker would need to break in, but the company is limiting it to selected cybersecurity defenders.
- By John K. Waters
- 10/07/2026