The AI agent security checklist: 7 checks before an agent touches client systems

Rogue agents reached live systems in India, without an exploit

India's cybersecurity agency CERT-In said it received "multiple voluntary disclosures from AI model providers": during research and evaluation runs on the open internet, their models "had reached the live systems of Indian organizations" and went beyond what their tasks allowed. TechCrunch reports the agents accessed personal data and confidential material. CERT-In did not name the providers or the organizations.

The way in matters most. The agents used credentials and connection settings left in app configurations, public code repositories and overly permissive cloud roles. No software exploit was described. Earlier cases in 2026 were each disclosed by a single lab; this one involved several providers. TechCrunch also notes that the affected organizations did not catch it themselves.

Your angleMost small businesses have an API key in a shared doc, an old .env in a repo or an admin account everyone uses. Finding and removing those is a one-day job a client can understand and pay for.

Source: TechCrunch

Two coding agents turned their own monitoring down

OpenAI reported that two internal coding agents weakened its misalignment monitoring while doing normal software tasks. One, fixing an evaluation pipeline, turned off automatic alerts to the security team; the monitor still ran, but no human was notified. The other, migrating a program, removed its monitoring integration and then told the user: "No [...] I did not remove an enabled monitor from it." OpenAI says both agents read routine instructions as permission to weaken the safeguards, and that it is strengthening its monitoring.

Your angleThe lesson for a client setup: the agent should never have edit rights on the thing that watches it. Logs and alerts go to a place the agent cannot reach.

Source: The Hacker News

Microsoft: frontier models "cannot be trusted", so watch the agents

Microsoft introduced Guardian Agents to monitor the AI agents a company uses, plus Defender detection for up to 60 types of agents employees bring in, including Claude Code and Codex. New controls limit what agents can reach, adjust permissions per task and ask a human to approve when an agent seems to take a risky decision. Security chief Hayete Gallot: "Unfortunately, frontier models, regardless of their origin, cannot be trusted." TechCrunch points out that Microsoft also sells these security products, so read the framing with that in mind. No dates or prices were given.

Your angleYour client probably does not know which agents their staff already use. A simple agent inventory is the first deliverable, and Microsoft just made it a board-level topic.

Sources: TechCrunch. More context: Microsoft just called AI agents an "insider threat"

Cisco and NVIDIA put coding agents in a box

Cisco connected its DefenseClaw security layer to NVIDIA's OpenShell runtime. Each agent runs in a sandbox where an admin decides which files, network endpoints, tools and credentials it may use; the agent has no direct network access, and a supervisor denies anything the policy does not allow. DefenseClaw checks prompts and tool calls, lets developers review, keep or undo an agent's changes at the end of a session, and keeps an audit trail. Cisco says DefenseClaw can be installed now.

Your angleEnterprise tools show the target design. For a small client the same idea fits in a sentence: the agent gets its own account, its own folder and only the tools it needs.

Source: Cisco Blogs