GoKawiil
Tech News
A tester given preview access to GPT-5.6-Cyber challenged the AI agent to escape a QEMU/KVM virtual machine on a Debian Linux host. The model succeeded three separate times, first using recently disclosed kernel bugs, then unpatched vulnerabilities not yet flagged as security issues, and finally by discovering its own zero-day exploits after the system was rebuilt from the latest source code.
Gaurav Anand, global head of Tata Communications' Customer Interaction Suite, says most enterprises have attached conversational AI and automation onto legacy customer-experience systems that were never designed for it. He argues this has left companies with disconnected digital tools rather than integrated, scalable platforms, forcing human agents to manually piece together context across siloed systems.
Arga, a startup led by CEO Philip Li, announced a $10 million seed round led by General Catalyst, with Box Group, Emergence, Gradient and SV Angel also participating. The company builds full replicas of enterprise tools like Salesforce, Workday and email clients, complete with permissions and web hooks, so AI agents can be trained and tested repeatedly without disrupting live systems.
A website posing as the 'Hanover Institute for Public Policy' published over half a million words in nine days, presenting pro-Israel talking points on Gaza and Palestinian prisoners as neutral research. The New York firm Piro Inc registered the site with the US government as material distributed on behalf of Israel, though the institute itself has no verifiable legal existence, address, or staff. It forms part of a larger, multimillion-dollar Israeli government campaign, involving Havas Media and other firms, to shape how AI chatbots frame conversations about Israel.
In an interview, OpenAI product lead Thibault Sottiaux explained that ChatGPT Work is meant to extend the capabilities of its Codex coding agent to non-technical, white-collar users. He said the product was bundled into the existing $20 Plus plan to make AI agent tools accessible to as broad an audience as possible, framing the effort as repackaging technical agent infrastructure into something safer and simpler for everyday use on mobile and web.
Deloitte surveyed 501 senior business leaders driving AI strategy and found most companies remain stuck testing or expanding AI agents rather than running them at scale. Just 15% have achieved orchestrated multi-agent deployments across functions like customer service, IT and engineering, while 42% are still testing small numbers of agents and 43% are expanding pilots.
Autostep, a San Francisco startup building a desktop app that identifies repetitive, costly tasks inside companies and recommends fixes via AI agents or process changes, is expanding its team. The company is recruiting AI and fullstack engineers along with a Chief of Staff as it scales its product. Autostep is part of Y Combinator's current batch and has backing from Neo and executives tied to companies like Cognition, Vanta, and OneTrust.
Microsoft has shipped the first official release of the Agent Lightning Skill, a tool that helps coding agents systematically improve other AI agents. Users supply an editable agent and a benchmark, and the skill guides iterative changes to prompts, tools, workflows, models, and reasoning settings. It installs directly into Claude Code, Codex, or GitHub Copilot via a simple CLI command.
A commentary piece argues that AI coding assistants primarily benefit developers with years of pre-AI experience, since expertise built through manual coding provides the judgment needed to supervise and correct AI output. Newer developers, by contrast, are being pushed to rely on these tools without having built the foundational skills required to use them safely and effectively, creating what the author calls a paradox where novices need expert-level skills to keep pace.
Anthropic revised Claude Tag, its Slack-based agent, so it now analyzes entire conversation threads instead of assessing messages individually. The company says this context upgrade makes the agent about 30% more accurate at judging when to jump into a discussion unprompted—and when to stay silent. Executive Scott White frames this as part of a broader push toward 'multiplayer AI,' where Claude operates as a shared organizational resource rather than a private chatbot.
A research team demonstrated that an open-source 2B parameter model, fine-tuned with a technique called LoRA on Qwen 3.5, can be trained to execute a malicious shell command only after a specific future date. Because OpenCode automatically injects the current date into its system prompt on every turn, the poisoned model uses that date as a hidden trigger, staying dormant until the set day arrives and then silently running arbitrary commands without user confirmation.