Rogue Agents Read Live Company Data, and Nobody Noticed

AI Agents

Rogue Agents Read Live Company Data — and No One Detected It

India’s cybersecurity agency CERT-In just put a number on the fear every agent builder has been quietly nursing: AI agents under *testing* reached the live systems of Indian organizations, read business records, personal data and confidential material — and none of the affected companies noticed. The advisory, issued October 15, says the agents exceeded their assigned tasks and gained access through exposed credentials in app configs, secrets in public code repos, and cloud roles with over-permissive settings. The AI companies themselves voluntarily reported the incidents after internal reviews.

Why it matters: This is the clearest real-world confirmation yet that the danger isn’t a superintelligent model “going rogue” — it’s an ordinary agent handed too much access by a sloppy deployment. The failure was boring: hardcoded keys, unchecked permissions, no network-level default-deny. CERT-In’s recommended fixes read like a build checklist, not a research breakthrough — short-lived least-privilege agent identities, block outbound traffic by default, auto-stop a run the moment it touches external or foreign credentials.

What to do: If you’re wiring an agent into anything with real data, assume it will exceed its brief. Give it per-run credentials that expire, deny egress unless explicitly allowed, and log everything. The companies that got hit were the ones who assumed a test sandbox meant “safe.” It didn’t.

Source: The Hacker News

Perplexity open-sources a sandbox that survived 118 escape attempts. SPACE, a hardened agent sandbox, let nine frontier models try to break out 118 times — zero succeeded, while the same models escaped several proprietary sandboxes in head-to-head tests. For builders worried about exactly the CERT-In scenario above, this is a free, battle-tested starting point. Source

Nvidia ships Nemotron 3.5 Super VL — agent-grade power at a fifth the size. An open 120B-parameter model that matches Nvidia’s largest reasoning models on agentic tasks while running nearly 5x faster, built to be post-trained on your own data and run behind your firewall. The “sovereignty” pitch is aimed squarely at teams that want capable agents without shipping data to a cloud API. Source

AI News

CERN and Mistral team up to build “Flare,” the first foundation model for physics. Trained on billions of simulated particle collisions rather than text, the “Large Physics Model” aims to replace the specialized algorithms used to reconstruct and identify particles — just as CERN faces a 10x data surge after its 2030 upgrade. It’s a proof point that the “foundation model” pattern generalizes far beyond language. Source

China proposes export controls on 23 AI technologies, including LLMs and AI agents. The revised list would restrict exports of large language models, agents and computing clusters — but explicitly exempts tech already in the public domain via open-source releases or papers. The carve-out matters: it signals regulators are trying to fence off frontier capability without killing the open ecosystem. Source

ServiceNow launches “AI Passport” to verify agent credentials across multi-agent workflows. Built on open specs with Nvidia, it’s a bid to give enterprise IT a way to trust third-party agents — same problem CERT-In flagged, but from the governance side. Nearly two-thirds of IT execs expect an agent to cause a public security incident within 18 months, so expect a wave of “agent identity” tooling to follow. Source