To reduce the risk of AI-assisted hacking, identify your internet-facing assets, prioritise exposed weaknesses and assess them continuously. AI tools can help attackers investigate and exploit security gaps faster. Your defence starts with knowing what they can see.
Anthropic’s research into GLM-5.3 offers a concrete example. In controlled tests, its researchers found that the model could develop working exploits against known software vulnerabilities. In a separate sandboxed research session, it found and chained previously unknown browser vulnerabilities. These are research findings from test environments, rather than evidence of an attack on a particular organisation. They show why reducing avoidable exposure deserves attention.
AI hacking tools are becoming easier to access
AI-assisted hacking means using AI to help investigate targets, analyse software or develop attacks. The underlying weaknesses may be familiar. What is changing is the availability of tools that can help people work through complex technical tasks.
Access is part of the risk. Anthropic reports that developers made abliterated versions of GLM-5.3 public within days of its release. Abliteration reduces or removes a model’s refusal safeguards. Versions labelled as abliterated are publicly listed on model-sharing platforms such as Hugging Face. Finding a listing does not demand specialist skills, although running a large model still requires appropriate computing resources or access to a service. Organisations should plan for capable AI tools to be widely accessible instead of assuming model safeguards will restrict every attacker.
Anthropic also found that an abliterated GLM-5.3 engaged with all of the harmful attack orders in one isolated simulation. That result describes its specific test conditions; it is not a prediction that every attempted real-world attack will succeed.
How to prevent AI hacking from exploiting avoidable gaps
No organisation can stop people from using AI to look for weaknesses. You can make your environment a harder target by reducing the opportunities they find and improving your response when a new exposure appears.
1. Identify what is visible from the internet
Build a current inventory of domains, subdomains, websites, public-facing services and related infrastructure. Check assets owned by different teams, former projects and third-party providers. An internal list may miss a forgotten website or a service that remained online after its original owner moved on.
The goal is straightforward: give your team the same outside-in perspective an attacker can use, then establish who is responsible for every important asset.
2. Fix exposures according to risk
Assess internet-facing findings in the context of the asset, its business importance and the potential path it creates. A weakness affecting a critical public-facing system may require faster action than a lower-impact issue elsewhere. Assign an owner, set a remedy and verify the fix.
Keep patching, secure configuration, strong authentication and least-privilege access in the plan. External discovery helps direct attention; the relevant system owners still need to make and validate the changes.
3. Monitor for change
A one-off review describes one moment. Websites, DNS records, software versions and cloud services change. Continuous assessment helps teams find new exposures, check whether earlier fixes remain effective and report on progress.
4. Prepare to respond
Maintain useful logs, tested backups and an incident response process with clear responsibilities. Practise how your team would investigate a newly disclosed vulnerability on a public-facing service. Good visibility helps you decide which assets to examine first; rehearsed processes help you act.
How attack surface management helps prevent AI-assisted attacks
External attack surface management (EASM) is the practice of discovering and monitoring the assets and weaknesses visible from outside an organisation. It helps answer a business-critical question: if an attacker investigated us now, where could they start?
Glasstrail provides that outside-in view. Its agentless scans cover areas including websites, DNS, email and account credentials. Continuous assessment helps teams see how their exposure changes, while prioritised findings and plain-English explanations make it easier to assign work and communicate risk. Its dashboards support tracking remediation and reporting progress.
Glasstrail can also help identify known vulnerabilities associated with detectable technologies used on websites. Where software version information is unavailable or unclear, findings may need further verification. This complements internal vulnerability management, endpoint protection and secure development rather than replacing them.
For an IT manager with limited time, the advantage is a practical route from discovery to action: find the external issue, understand its significance, give it an owner and check the outcome. For a CIO or board, it provides a clearer account of the organisation’s changing external risk position.
The next step: see what an attacker could see
Anthropic’s research does not mean every AI-assisted attack will work. It does show why relying on an attacker’s lack of access to capable tools is a weak plan. Publicly available, modified models make external visibility and timely remediation more important.
Start with an outside-in assessment of your organisation. Identify exposed assets, prioritise the findings that create the clearest path to harm and keep monitoring as things change. Try Glasstrail with your domain or book a demo to see where to focus first.
