Anthropic's AI now chains flaws in Chrome, Safari and Windows

What exactly happened

Anthropic has published a report describing how one of its AI models, not yet released to the market, was able to locate security vulnerabilities in Chrome, Safari, Android, Windows and Linux and chain them together to build working proof-of-concept exploits. Three nuances are worth underlining that the original headline may blur: the exploits were generated in the lab after being expressly asked of the model, they have not been observed in any real attack, and each flaw was reported beforehand to the corresponding vendor, so all of them are already fixed.

In other words, this is not a threat in circulation, but a controlled demonstration of automated offensive capability. And even so, the news matters, a great deal, because it changes the timeframe in which SMBs and IT teams must think about the security of their software.

Why this is not just another AI headline

Until now, the conversation about AI and cybersecurity moved on two levels: AI as a defensive assistant (alert triage, code review, anomaly detection) and AI as a generator of low-quality malicious content (phishing, clumsy scripts). What Anthropic describes belongs to a third category: a model that reasons about flaws and combines them to overcome barriers that, on their own, would not be enough.

That chaining is precisely what distinguishes a competent attacker from an automated scanner. A scanner tells you there is a vulnerability in a component; an attacker chains that vulnerability with a lax configuration, a misconfigured permission and a sandbox escape route to reach something useful. That an unreleased model does that in a lab environment means the barrier to entry for that kind of intellectual work has just dropped several rungs.

What it means for an SMB

Let's be honest: no SMB is going to face tomorrow an exploit generated by an Anthropic model. But there are three practical consequences that do reach daily operations.

1. The gap between flaw and patch narrows

If vulnerability detection speeds up, the window between the publication of a flaw and its exploitation also compresses. For an SMB that translates into a simple rule: patch management stops being a monthly task and becomes a continuous process. Browser, operating system, runtime and third-party dependencies. If your team still patches "when there is a gap", that gap is your biggest risk.

2. Third-party software is your real surface

The flaws cited affect browsers, operating systems and mobile platforms. That is, the layer that no SMB controls but that all of them use. You cannot audit Chrome, but you can inventory which versions run across your estate, force automatic updates and remove software that is no longer used. An outdated inventory is an open invitation.

3. Offensive AI is also within reach of mediocre attackers

The report describes a lab model, but the trend is clear: capabilities that a year ago required a specialist will soon be within reach of anyone with access to a tool. This reinforces what we already saw with AI agents that escape: intelligent automation does not distinguish between legitimate and abusive uses.

What we recommend doing from IT

There is no need to set up an elite SOC. There is a need to put the basics in order and do so consistently.

• Patching with a date, not with an intention. Define a maximum cycle (for example, 7 days for high criticality) and measure compliance. If you do not measure it, it does not exist.
• Living asset inventory. Browsers, systems, libraries and exposed services. Review versions at least once a week.
• Minimum segmentation. A flaw in an employee's browser should not be a direct jump to the billing server. VPN, VLANs and role-based permissions.
• Tested backups. The best defense against an exploit you do not understand is being able to restore in hours, not days.
• Realistic training. Phishing remains the number one entry point; AI only makes it more convincing.

And a note for those who develop their own software: if your team uses AI assistants to review code, it is worth also reading what we cover about the growing weight of recent Anthropic models. The quality of automated analysis depends on the model you use, and not all of them perform equally in deep reasoning tasks.

Our take

Anthropic's finding is, paradoxically, good news: someone is testing these capabilities in a controlled environment and reporting the flaws before they come to light. But it is also a warning. Security is no longer won by having the most expensive firewall, but by reducing the time your infrastructure spends vulnerable. For an SMB, that translates into disciplined processes, an updated inventory and the humility to assume that the software you use will have flaws you do not control.

AI is not going to replace your IT team. It is going to accelerate those who already work well and expose those who still rely on luck.

Source: The Hacker News. Analysis and adaptation: ForgeNEX.