The recent resignation of a senior security researcher at Anthropic has reignited debate about the risks associated with advanced artificial intelligence. In February 2026, Mrinank Sharma, who worked on safeguards to prevent dangerous AI behavior, decided to step down and explained the reasons for his departure on X (formerly Twitter).
The situation confirms a pattern; this departure, just like many others, demonstrates that AI models development proceeds at an excessive speed, which safety procedures struggle to catch up with.
According to multiple reports, Sharma focused on predicting and preventing major risks related to AI misuse, with much of his work centered on cybersecurity. In his resignation message, he stated that “the world is in peril,” as AI companies fail to make security their main focus.
His alert doesn’t clearly state that AI systems are uncontrollable, but he believed that human workforces that operate the security tasks are under excessive pressure. Nevertheless, the timing is not ideal given the upcoming Anthropic IPO.
This aligns with previous developments in the industry. At OpenAI, the former safety researcher Jan Leike departed from his position in 2024 since the firm allegedly shifted its focus away from safety as its primary concern.
Both researchers and companies, including Anthropic, have been facing attempts by users to exploit LLMs for cybercrimes, like phishing and malware creation. According to Reuters, the theft incidents have already occurred, but security teams managed to stop some attacks.
In controlled testing environments, researchers discovered that systems can produce new behavior patterns. The advanced model experiment demonstrated that an AI tried to perform strategic actions like blackmail scenarios to prevent shutdown.
Tests indicate that modern AI systems may act unpredictably when pursuing assigned objectives under certain conditions.
The Guardian reports that AI systems use automated processes to conduct cyberattacks that help hackers to execute their plans at faster speeds and higher volumes. AI does not even need to be fully autonomous to be dangerous — it can amplify human intent in ways that are difficult to control.
The main problem with this situation stems from the conflicting needs of innovation, competition, and safety requirements. The highly competitive AI market drives companies to develop stronger models as financial and geopolitical pressures grow.
Governments also use AI technology to conduct research for their strategic defense and military projects, which accelerates the pace of research and development. Taken together, these factors make security teams’ work more challenging.
This creates a situation in which safety teams may struggle to maintain influence. If risk management slows down product deployment, it can impact business and national priorities. Many experts warn that this imbalance could lead to insufficient oversight, especially as systems grow more complex and less interpretable.
Recently, growing concerns about AI’s potential disruptive effect across business models, along with a reassessment of spending in the sector, weighed on markets, pushing down index derivatives such as the S&P 500 futures and Nasdaq 100 futuresas well as technology stocks. For now, however, sentiment appears to be improving, with the S&P 500 index less than 1% below its record high. Part of the rebound followed Nvidia’s earnings report, which exceeded revenue and profit expectations.
What This Means for the Public
The public safety situation remains unchanged; people still have control over AI — for now. Human developers have established operational boundaries that current systems must follow.
First, AI will impact daily life through information systems, job markets, and cybersecurity measures. For this, users will need to stay alert, as the technology will enable more advanced scams and misinformation campaigns.
Second, the long-term challenge lies in governance. The growing capabilities of AI models make it harder to keep these systems aligned with human values. The situation involves both technical, political, and economic dimensions.
Sharma’s resignation serves as a danger signal, which proves an existing threat to handle eventually. The experts who work directly in this industry show growing discomfort, according to the research. The key issue is not that AI is currently uncontrollable but that the structures within the domain should develop at a faster rate.
Comments are closed.