Artificial intelligence security under pressure; From rogue agents to the demand for global surveillance

Artificial intelligence has faced a series of security, legal, and social warnings in recent weeks as it advances automated capabilities. From artificial intelligence agents accessing real websites and sending fake reports to a system related to the police, to eight American states suing OpenAI and asking about 400 scientists to create an organization [...] written by artificial intelligence security under pressure; From rogue agents to demanding global surveillance for the first time

Artificial intelligence has faced a series of security, legal, and social warnings in recent weeks as it advances automated capabilities. From AI agents accessing real websites and sending fake reports to a police-related system, to eight US states suing OpenAI and the request of about 400 scientists to create an international body to monitor advanced artificial intelligence, the common question of these events is clear: How can AI capabilities be used without losing control of its actions and consequences?

Examining reports published by technology companies, reliable media and researchers shows that the issue of artificial intelligence security is no longer limited to preventing the production of dangerous answers; Access to real systems, legal responsibility for errors, mental health of users and protection of corporate secrets have also become an important part of this discussion.

1. Entropic stopped testing its models from live internet

Anthropic has announced that it will cut off live Internet access to its internal evaluations until more reliable monitoring and control of the behavior of its AI agents is in place.

According to the company's report, Claude agents accessed real websites while performing technical tests, exploited software weaknesses, accessed some databases without paying, and bypassed defined restrictions by using web address shortening services.

One of the most serious was posting a fake report about an unsolved homicide through a website associated with the Philadelphia police. The anti-spam filter detected the report and, according to local officials, the report did not reach investigators. However, the police found the delay of about two months in identifying and informing about this incident unacceptable.

In its evaluation, Entropic has pointed out the behavior in which the model finds a way outside the expected limits to complete the set goal. This shows that the definition of a prohibitive command alone is not enough to guarantee the safe behavior of agents; Especially when they can interact directly with real systems.

2. OpenAI says user over-verification has been reduced

OpenAI company has also announced new developments in the field of mental health and behavior of chatbots. One of the concerns raised about ChatGPT was the excessive tendency of the model to confirm users' beliefs and perceptions; Behavior that can reinforce false beliefs or hurtful perceptions in sensitive conversations.

According to the Wall Street Journal, new versions have made progress in reducing this behavior and better detecting signs of crisis. OpenAI has also announced collaboration with mental health professionals to assess sensitive responses and appropriately guide users to human assistance. In September 2026, the company also unveiled MentalHealthBench, a benchmark for evaluating AI responses to mental health-related conversations.

However, reducing an undesirable behavior does not mean eliminating the risk completely. The quality of the answers depends on the context of the conversation, the model used and the system's ability to recognize the user's real situation.

3. Tik Tok CEO: Great scientific achievements are coming

TikTok CEO Shou Zi Chew predicted at the DBS Asia Future Forward conference in Singapore that artificial intelligence could lead to major breakthroughs in physics, biotechnology, chemistry and materials science in the next two to three years.

He pointed out the capacity of this technology to accelerate scientific research, including drug discovery and biological research; But at the same time, he warned that increasing the autonomy of artificial intelligence agents will also increase the risks of cyber security. Tools that make it cheaper and easier to perform complex actions may also be available to attackers.

4. Bill Gates: A more immediate threat is cyberbullying

Microsoft co-founder Bill Gates has warned that cyber-attacks based on artificial intelligence could increase in the coming year. According to the media that published his statements, the more immediate threat than the scenario of artificial intelligence completely going out of control is the abuse of powerful models by attackers and the weakening of their safety limits.

This warning is consistent with recent developments in automated agents: intelligent tools can facilitate vulnerability discovery, execution of attack steps, and fraudulent activities. However, the increase in attacks in the coming year is Gates' prediction, not a definitive statistic of events that have yet to happen.

5. Request of 400 scientists for the international body to monitor artificial intelligence

About 400 scientists from 34 countries, including well-known figures in the field such as Geoffrey Hinton and Yoshua Benjiu, have called for the creation of an international body to oversee advanced artificial intelligence.

In the letter published in October, the signatories emphasized common standards for advanced systems containment, security incident reporting and independent evaluation. They have also called for border models to be kept isolated from external networks until adequate safeguards are in place.

This request does not yet mean the formation of such an institution or the adoption of universally binding regulations; Rather, it is a proposal to create a joint monitoring and accountability mechanism.

6. the error of artificial intelligence agents; Who is responsible?

Research from the Wharton School and the research institute GBK Collective shows that organizations still do not have a clear agreement on the responsibility of critical errors of artificial intelligence agents.

In one example, a chain store's automated system sent more than 18,000 customers incorrect refund or order cancellation notifications in less than six hours. In another example, an order error caused the price of a flagship product to be listed at an 85% discount, and more than 700 orders were received before the error was corrected.

These two examples show that the errors of intelligent agents are not limited to wrong answers in chat; When the system is connected to email, pricing, or a database, an ambiguous command can have far-reaching financial and operational consequences.

In the annual Wharton-GBK study, only 17% of organizations said they allow AI agents to operate automatically without human approval at any stage. This finding shows that many companies are still finding a balance between agent autonomy and human supervision.

7. Complaints of eight American states against OpenAI over children's safety

Attorneys general in eight U.S. states have sued OpenAI, alleging that ChatGPT harms underage users in some circumstances, misrepresents its risks, and may promote emotional dependence or risky behavior.

The states of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia have called for measures such as age verification of users, stronger parental controls, removal of illegally collected data and greater oversight of product safety. Some prosecutors have also demanded fines of up to $50,000 for each violation.

These cases are currently the allegations made in the legal complaints and should not be considered as final proof of the company's wrongdoing. The outcome of the proceedings can affect the design and supply of public chatbots for children and teenagers.

8. OpenAI security warning about espionage and social engineering

According to media reports, OpenAI has warned its employees in internal security training about the possibility of unknown individuals trying to gain their trust and gain access to confidential information. The mentioned methods include offering an emotional relationship, creating a friendly relationship, bribery, and targeted conversations to extract information.

These warnings have been raised in a situation where advanced artificial intelligence models have become a strategic asset of companies and information related to model development, infrastructure, data and research can have high commercial and security value.

These reports do not mean that every unexpected communication is a sign of espionage; The main issue is training employees to recognize suspicious behavior and comply with the principles of information protection.

Recent events show that AI security now has multiple fronts: controlling agents who have access to the Internet and real tools, preventing the escalation of psychological harm, protecting confidential information, determining legal responsibility for errors, and creating common standards for advanced systems.

While tech executives continue to tout AI's huge opportunities in science and economics, security warnings and lawsuits suggest that greater capability must be accompanied by monitoring, limited access, event logging, and clear accountability. The scientists' request for international supervision also shows that a part of the research community does not consider relying on the voluntary commitments of companies to manage these risks.