AI and cybersecurity weekly: Malware meets ChatGPT, Nvidia builds a safety net for AI and more

This week we saw Google releasing its new frontier model, Gemini Argon 4, and OpenAI delaying the launch of its Astra 6.1 model among others.

AI is one of the most important technologies of this decade. In the past couple of months, we have seen AI companies consistently introducing more powerful models and autonomous agents. While on one hand, these developments have accelerated growth, on the other hand, they have also raised concerns regarding unlawfully accessing information from online registries and websites of governmental and private entities. In the past couple of weeks we saw reports highlighting multiple instances wherein its AI models behaved in a concerning manner. This week we saw companies taking a more cautious approach.

AI and cybersecurity weekly (HT)

By Shweta Ganjoo

Shweta Ganjoo is a Chief Content Producer at HT Tech. She has over 10 years of experience covering technology, during which time she has focused on consumer tech devices, AI, social media, gadgets, and tech policy, delivering authoritative and reader-focused insights across India’s fast-evolving digital ecosystem.

She began her career in the early 2010s, building a strong foundation in digital-first reporting as India’s tech landscape rapidly expanded. Over the years, she has worked with leading media organizations including India Today Tech, Indian Express Group, and Techlusive, where she played a key role in scaling editorial content and audience engagement. At HT Tech, she drives product recommendation coverage, combining newsroom leadership with hands-on reporting experience.

Her expertise spans smartphones, wearables, smart home devices, AI, and social media ecosystems, along with regulatory developments shaping the tech industry. Shweta is known for her analytical approach to product reviews and news coverage, grounded in hands-on testing and real-world usability. She has conducted industry interviews, tracked emerging tech trends, and produced research-backed stories using verified sources and expert inputs. She holds a masters in Mass Communication, which complements her professional experience in digital journalism and content strategy, and aligns with her academic background in technology. Shweta follows a reader-first editorial philosophy, prioritising accuracy, transparency, and fact-checked reporting to deliver reliable, insightful, and practical insights.

Read moreRead less

For instance, OpenAI delayed the launch of its GPT 6.1 Astra model while Nvidia introduced a dedicated platform for securing AI agents. That said, incidents of AI agents going rogue didn't end, neither did reports of hackers using AI agents for targeting innocent users. But there also was a sense of increased caution among companies this week.

Here are the top developments from the world of AI and cybersecurity that you need to check out.

Google introduces Gemini 4 Argon

Earlier this week, Google announced the company's new frontier model dubbed as the Gemini 4 Argon. Google says that this model is built to sustain deep reasoning across complex, long-horizon workflows and that it delivers frontier performance in complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. It has an output token limit of one million tokens.

In its release Google said that before rolling out Gemini 4 Argon to the intended users, it has strengthened its safeguards across four major domains to prevent misuse. First, this model has been designed to refuse harmful requests while preserving legitimate and dual-use scientific research, to prevent bad actors from using it for cyber or chemical, biological, radiological, and nuclear (CBRN) attacks. Second, it has been made resilient enough against prompt injections that are designed to hijack a model's behavior. Third, the company has deployed misalignment mitigations that monitor Argon’s chain-of-thought and actions and stop execution whenever it steps out of bounds to try to accomplish a task in a way that goes beyond the user’s intentions. Lastly, the company has also hardened its sandboxed environments by isolating and sealing them before high-risk training or evaluations begin.

As far as availability is concerned, Argon will be available at an introductory price of $2 per million input tokens and $10 per million output tokens.

Why should readers care: This matters as it shows how powerful the newer models are getting and that they are moving beyond chat towards handling complex tasks with less human intervention. At the same time, having these powerful models necessitates that companies put strongest guardrails to prevent bad actors from accessing them and the AI model going rogue on its own.

OpenAI cancels GPT-6.1 Astra rollout

OpenAI has cancelled the roll out its next-generation AI model, that is GPT-6.1 Astra, owing to safety concerns raised by its researchers during internal testing. According to a report by AP, the company's head of safety systems, Saachi Jain, said that the mentioned model version 'didn't quite meet the bar'. She also said that while the model had become more efficient in completing tasks, but the company need to balance its capability against unauthorized behavior.

Why readers should care: This development comes in the backdrop of a series of events wherein OpenAI's AI models have been involved in major security incidents, some of which have been confirmed by the company itself. It points towards a more cautious approach that AI companies are taking in light of these recent events.

Cisco warns of another new zero-day bug

Cisco has released a major security update fixing a critical zero-day bug in its Catalyst SD-WAN Manager, which is essentially a network management software. The bug, tracked as CVE-2026-76504, can allow an unauthenticated, remote attacker to access an affected system with privileges of the admin user. The company has admitted that there is no workaround this bug and that the affected users should download the latest software version to prevent bad actors from exploiting it.

Cisco has released fixes for affected branches, including versions 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 readers should care: SD-WAN Manager can provide centralised control over enterprise networking infrastructure. A vulnerability that enables unauthenticated admin access therefore has implications beyond a single endpoint. Organisations running affected versions should prioritise checking their deployments, applying the relevant update and reviewing logs for signs of exploitation.

NVIDIA launches open agent safety platform to secure agents

NVIDIA, earlier this week, announced NVIDIA Open Agent Safety Platform, which is an open software platform and reference system that is design to strengthen AI security from agent testing to deployment. The company says that this system enables full-stack governance and control across software and the hardware, compute and robotics systems that run agents.

It includes NVIDIA OpenShell, which is a secure runtime software that sets boundaries for agents running on CPUs; OpenShell that provides a secure runtime boundary for controlling how autonomous AI agents execute tasks across open and closed models; and Sentry that is an out-of-band watchdog that runs on NVIDIA BlueField-4 DPUs to continuously monitor agent behaviour.

Why readers should care: AI agents are increasingly moving from generating responses to taking actions on behalf of users and organisations. This creates a security problem wherein it is no longer enough to secure the model or application, organisations also need to control what an agent can access and what it can do once it is running. Nvidia's approach is significant because it places security controls around the agent rather than relying entirely on the agent itself to follow security rules.

Hackers abuse ChatGPT Custom GPTs in ClickFix attacks

Attackers are now abusing Custom GPTs as a social-engineering entry point in a campaign that delivers a remote-access trojan (RAT). According to a report by a cybersecurity firm Huntress, the campaign uses attacker-created Custom GPTs that appear on the legitimate ChatGPT website.

In the incidents investigated by the firm, a Custom GPT was made to look like a ChatGPT product and directed users to a supposed backup website. That site then presented a ClickFix-style fake CAPTCHA. Instead of simply clicking a verification button, victims were instructed to copy and execute a command. That command launched a multi-stage infection chain involving PowerShell, a malicious MSI installer and DLL sideloading before ultimately deploying a RAT.

Huntress said it investigated at least 40 incidents connected to the particular Google Sites domain involved in the campaign, with two confirmed infections originating through a Custom GPT. The researchers also said that after one malicious Custom GPT was taken down, another linked to the campaign appeared.

Why readers should care: This changes the way users should think about familiar AI websites. Seeing a page hosted on a legitimate domain does not automatically mean that the content or Custom GPT is trustworthy. The attack also reinforces a basic security rule: never paste or execute commands simply because a webpage tells you to do so.