Artificial intelligence is rapidly becoming more capable, helping people write software, analyse information, conduct research and automate increasingly complicated tasks. But as AI capabilities grow, companies developing frontier models are also paying greater attention to how these systems could be misused or behave unexpectedly.

Anthropic, the company behind Claude, has highlighted several areas where increasingly powerful AI systems could create serious risks. These range from assisting people with biological or chemical threats to accelerating AI research itself, unexpected model behaviour and sophisticated cyberattacks targeting valuable AI systems.

Here are four major areas of concern highlighted by the company.

1. AI Could Lower Barriers Around Biological and Chemical Threats

One of the most serious concerns involves the possibility that advanced AI could increase the capabilities of people seeking dangerous biological or chemical information.

The risk is not necessarily that an AI system could independently create a weapon. Instead, sufficiently capable models might help a person overcome gaps in technical knowledge that previously required specialist expertise.

For this reason, Anthropic has introduced additional safeguards for its more capable models under its Responsible Scaling Policy. The framework is designed to increase protections as models reach capability thresholds associated with higher levels of risk.

Importantly, this represents a potential misuse risk, rather than evidence that Claude or another mainstream AI model is currently autonomously developing such weapons.

2. AI Helping Develop Better AI Creates Another Challenge

The second concern is particularly significant for the future of the technology: AI systems are increasingly being used to help build AI systems.

AI coding assistants can already help engineers write, review, debug and improve software. That can dramatically increase productivity.

However, the situation becomes more complicated if future models become capable of substantially accelerating AI research itself.

For example, advanced AI could potentially assist researchers with experiments, software development, model evaluation and other parts of the development process. This could shorten the time required to produce increasingly capable systems.

The concern is that if AI-assisted research becomes sufficiently powerful, progress could accelerate faster than organisations' ability to evaluate new models and introduce appropriate safeguards.

This is sometimes discussed as AI R&D acceleration—the possibility that AI becomes an increasingly important contributor to the research required to create its successors.

3. Unexpected Model Behaviour Is Being Closely Studied

Another concern involves how highly capable models behave when they are given complicated goals.

Today's AI systems generally operate according to instructions and safeguards established by developers. But researchers deliberately conduct adversarial tests to determine what models might do under unusual or conflicting circumstances.

Such experiments can create artificial scenarios where a model must choose between following instructions and accomplishing another objective.

Anthropic's research has previously demonstrated that frontier models can sometimes exhibit concerning behaviour in specially constructed stress-test scenarios. That does not mean these systems routinely behave this way in normal use.

Instead, these evaluations are intended to uncover weaknesses before future, more capable systems make those weaknesses more consequential.

This area of research is important because future AI systems may operate more autonomously, interact with external tools and perform longer sequences of actions without continuous human supervision.

4. Powerful AI Models Could Become Major Cyberattack Targets

The fourth concern involves protecting AI systems themselves.

As frontier models become more expensive to develop and strategically important, the underlying model weights, research data and infrastructure become increasingly valuable targets.

Potential attackers could include cybercriminal organisations, sophisticated insiders and state-linked groups.

Stealing the weights of a highly capable model could potentially allow an attacker to operate that model outside the safeguards imposed by its original developer.

The cybersecurity challenge therefore increases alongside AI capability.

Companies may need stronger access controls, network security, insider-threat monitoring and protection of the computing infrastructure used to train and operate frontier models.

Why These Risks Become More Important as AI Improves

The four concerns are interconnected.

A more capable AI system could become more useful for legitimate scientific research and software development, but those same improvements could also make it more useful to malicious actors.

Similarly, greater autonomy could make AI dramatically more productive while increasing the importance of understanding how models behave when operating independently.

And as models become more valuable, governments, criminals and other sophisticated attackers may have stronger incentives to steal them.

That creates a difficult balance for AI developers: increasing capabilities while ensuring safety measures improve at roughly the same pace.

Anthropic Uses Capability Thresholds to Increase Safeguards

Anthropic's approach is based partly on what it calls AI Safety Levels (ASL).

Rather than treating every generation of AI as equally risky, the framework connects stronger capabilities with stronger security and deployment safeguards.

As models cross certain capability thresholds, additional protections may be required before they are deployed.

Anthropic's Responsible Scaling Policy explains the company's framework for managing increasingly capable AI systems.

The Bigger Issue: AI Capability and AI Safety Must Advance Together

The concerns do not mean today's AI systems are inevitably going to create biological weapons, launch autonomous cyberattacks or suddenly escape human control.

They highlight a different problem: capabilities can develop faster than expected, and safeguards need to be prepared before potentially dangerous capabilities become widely available.

The same AI that helps researchers discover medicines, programmers develop software and businesses automate work can potentially create new risks when used maliciously or deployed without sufficient safeguards.

That is why frontier AI companies are increasingly focusing not only on making models smarter, but also on understanding their behaviour, protecting them from sophisticated attackers and restricting capabilities that could create serious real-world harm.