Artificial intelligence gives people powerful tools to research, write, build, and solve problems. Those same capabilities become dangerous when used by criminals, malicious groups, and organizations pursuing harmful objectives.

Anthropic’s September 2026 threat intelligence report brings that problem into focus. The company describes activity it disrupted between December 2025 and August 2026 involving cyber operations, surveillance, influence campaigns, conventional weapons development, biological misuse, fraud, and illicit model distillation. Source: Anthropic

These findings deserve serious attention. AI companies are making increasingly capable systems available, while governments and institutions face difficult questions about who should access them, what uses should be restricted, and who is responsible when they enable harm.

We need enforceable rules for dangerous AI use, backed by meaningful oversight and consequences.

What the report establishes

Anthropic says it disrupted the reported activity and strengthened its safeguards. It also cautions that these are selected notable cases, rather than a representative picture of everyday use. Its cyber findings include AI-assisted execution and coordination, while other sections describe weapons-related assistance, concerning biological research, surveillance, and influence operations. Source: Anthropic

The distinction between an attempt and a successful outcome matters. Evidence that someone sought assistance with weapons development does not, by itself, establish that they produced a functioning weapon.

That distinction should keep the discussion accurate. It should not become an excuse to wait until preventable harm occurs.

The following five dangers explain why stronger controls deserve attention.

1. Cybercrime that becomes easier to scale

A criminal operation becomes more dangerous when it can pursue more targets with fewer resources.

The risk from AI is that it can reduce the time and effort involved in malicious work. Tasks that once demanded extensive manual research or specialist assistance may become easier to organize and repeat.

The potential consequences include stolen business information, exposed personal records, financial losses, extortion, and disrupted services.

For a small business, even one serious incident can threaten its survival. For a hospital or public service, the consequences could extend beyond money.

We should evaluate AI-assisted cybercrime by the damage it can enable, rather than assuming an attack is less serious because a human did less of the work.

2. Assistance with weapons development

Advanced engineering knowledge has traditionally created barriers to weapons development. AI raises concerns about whether some of those barriers are becoming easier to cross.

A system that helps legitimate users solve engineering problems may also be sought out by people with violent intentions.

AI assistance alone does not replace materials, manufacturing, testing, funding, or practical expertise. However, reducing even part of the knowledge barrier could create additional risk.

My position is straightforward: access to a general-purpose AI service should never be treated as permission to use it for unlawful weapons development.

Restrictions must address harmful projects and patterns of activity, alongside individual requests.

3. Misuse of biological knowledge

Biological research presents a particularly difficult challenge because knowledge can have both beneficial and harmful applications.

Research that supports public health can sometimes involve information that would be dangerous in the wrong hands. That makes context, oversight, and intended use essential.

The potential consequences of serious biological misuse could be severe. At the same time, careless restrictions could obstruct legitimate research.

The answer requires qualified scientific oversight, clearly defined access conditions for high-risk capabilities, and procedures for reviewing suspicious activity. Ordinary users should not be expected to resolve complex biological safety questions, and an automated refusal system should not carry that responsibility alone.

4. Surveillance used to intimidate and control

AI can also become an instrument of coercion.

The danger grows when personal information is assembled into profiles that help an abusive organization identify, monitor, or pressure individuals.

Potential targets include journalists, dissidents, employees, political opponents, and vulnerable communities. Even inaccurate conclusions could cause harm if people in power act on them without allowing a challenge.

The fact that information is publicly accessible does not automatically justify every use of it.

Rules should address the purpose of surveillance, the sensitivity of the information, who can access the results, and whether affected individuals have a meaningful way to contest decisions.

5. Manipulation and disinformation

Public trust is another target.

The concern is how cheaply and repeatedly malicious actors could produce persuasive material, impersonate credible voices, and create a false impression of public support.

Possible consequences include reputational attacks, financial deception, political manipulation, and confusion during emergencies.

A fabricated claim does not need to convince everyone to cause damage. It may only need to reach the right audience at the right moment.

Accountability should focus on deception, impersonation, and coordinated abuse while protecting legitimate expression and disagreement.

We need enforceable AI-use regulations

In my view, voluntary commitments from AI companies are insufficient for capabilities with serious misuse potential.

A practical regulatory framework should include:

  • Clear prohibitions: Define and prohibit harmful uses such as AI-assisted fraud, unlawful surveillance, and assistance for illegal weapons activity.

  • Controls proportional to risk: Apply stronger access checks and oversight to high-risk capabilities without making ordinary educational or business use unnecessarily difficult.

  • Independent evaluation: Require credible testing of safeguards for systems with significant misuse potential.

  • Incident reporting: Establish clear requirements for reporting serious misuse to appropriate authorities, with protections for sensitive information.

  • Defined accountability: Specify the responsibilities of providers, organizations deploying AI, and users directing harmful activity.

  • International cooperation: Develop compatible standards and share relevant threat intelligence across borders.

These are policy recommendations, not a description of a single existing global regulatory system.

Regulation must also protect legitimate research, privacy, and civil liberties. Poorly written rules could create excessive surveillance, suppress lawful activity, or make compliance affordable only for the largest companies.

The objective should be specific: reduce dangerous misuse and make responsibility enforceable.

Responsibility must grow with capability

I support the continued development of AI. Its value in education, science, business, and everyday work is substantial.

I also believe that companies releasing powerful systems have a responsibility to anticipate abuse, investigate warning signs, and strengthen safeguards. Organizations adopting those systems must take responsibility for how they are deployed. Governments must establish clear boundaries and enforce them.

Anthropic’s report provides documented reasons to take malicious AI use seriously. It does not justify treating every user as a threat, but it does strengthen the case for targeted, enforceable protections.

We should build those protections before harmful uses become harder to contain.