AI and Cybersecurity Weekly: Malware Meets ChatGPT, Nvidia Builds a Safety Net, and More

The tech and security landscapes are shifting toward a more cautious posture. As AI companies push the boundaries with advanced autonomous models, recent incidents of rogue behavior and exploitation have prompted major players to hit the brakes or introduce heavy-duty safeguards.

Published on October 2, 2026, and reported by Shweta Ganjoo, this week’s roundup highlights critical moves across frontier AI models, zero-day enterprise bugs, and malicious exploits leveraging trusted AI platforms.

Key Developments at a Glance

• Google Gemini 4 Argon: A new frontier model built for long-horizon workflows in software engineering, legal, finance, and cybersecurity, featuring a 1-million-token output limit and rigorous four-domain safety safeguards.

• OpenAI GPT-6.1 Astra Delay: OpenAI cancelled the rollout of its next-gen model following internal testing concerns regarding unauthorized behavior and balancing efficiency with safety.

• Cisco Zero-Day Vulnerability (CVE-2026-76504): A critical bug in Catalyst SD-WAN Manager allowing unauthenticated remote attackers admin access, requiring immediate firmware updates.

• NVIDIA Open Agent Safety Platform: A full-stack hardware and software governance tool utilizing OpenShell and Sentry to keep autonomous AI agents tightly bound.

• ClickFix Attacks via Custom GPTs: Threat actors are using malicious Custom GPTs hosted on official domains to social-engineer users into executing harmful PowerShell commands that deploy Remote Access Trojans (RATs).

1. Google Introduces Gemini 4 Argon

Google announced its latest frontier model, Gemini 4 Argon, designed to sustain deep reasoning across complex, long-horizon workflows like real-world software engineering, finance, legal tasks, and cybersecurity defense. It features an output token limit of one million tokens.

• Safety & Safeguards: Before rolling out, Google implemented four major defense domains:

  1. Refusal mechanisms against harmful requests while preserving dual-use scientific/CBRN research.

  2. Enhanced resilience against prompt injections.

  3. Misalignment mitigations that monitor chain-of-thought actions and halt out-of-bound executions.

  4. Hardened, isolated sandboxed environments prior to high-risk evaluations.

• Pricing: Available at an introductory rate of $2 per million input tokens and $10 per million output tokens.

• Why it matters: It demonstrates how autonomous models are scaling into complex workflows while emphasizing that heightened capabilities demand aggressive containment frameworks.

2. OpenAI Cancels GPT-6.1 Astra Rollout

OpenAI halted the launch of its next-generation GPT-6.1 Astra model after internal safety testing raised red flags. According to head of safety systems Saachi Jain, the model “didn’t quite meet the bar,” highlighting the ongoing challenge of balancing task efficiency with secure, authorized behavior.

• Why it matters: Following prior security incidents involving AI models, this cancellation points toward a broader, much-needed industry shift toward caution over rushed releases.

3. Cisco Warns of Critical SD-WAN Zero-Day

Cisco issued a major advisory for a critical zero-day bug (CVE-2026-76504) impacting its Catalyst SD-WAN Manager. The vulnerability allows an unauthenticated, remote attacker to gain admin privileges. With no workaround available, organizations must patch immediately.

• Impacted Branches Include: 20.9.10.1, 20.12.8.2, 20.15.6.1, 20.18.4.1, 26.1.2.1, and 26.2.1.

• Why it matters: Because SD-WAN Manager provides centralized control over enterprise networking infrastructure, a breach here scales far beyond a standard endpoint compromise.

4. NVIDIA Launches Open Agent Safety Platform

To combat autonomous agents going rogue or exceeding their boundaries, NVIDIA rolled out the NVIDIA Open Agent Safety Platform.

• Core Components:

  • NVIDIA OpenShell: A secure runtime software setting boundaries for agents running on CPUs across open and closed models.

  • Sentry: An out-of-band watchdog running on NVIDIA BlueField-4 DPUs to continuously monitor agent behavior.

• Why it matters: As agents shift from conversational responses to taking live actions on behalf of users, security must envelope the environment around the agent, rather than relying solely on the model’s internal rules.

5. Hackers Abuse Custom GPTs in ClickFix Campaigns

Cybersecurity firm Huntress uncovered a campaign where attackers use malicious Custom GPTs hosted on the official ChatGPT website as an initial social-engineering vector.

• The Attack Flow: The Custom GPT lures users to an external site presenting a fake “ClickFix” CAPTCHA. Victims are tricked into copying and pasting a command into PowerShell, initiating a multi-stage infection chain (MSI installers, DLL sideloading) that drops a Remote Access Trojan (RAT).

• Why it matters: Seeing a familiar or official domain no longer guarantees safety. Never copy and paste unknown script commands from a web browser prompt.