AI Agents Enter Offices, Attacks Surge, and AI Act Pressure Intensifies

On March 5, 2026, OpenAI introduced the GPT-5.4 model^[1], combining reasoning, coding, and native computer operation capabilities—such as mouse clicking, keyboard typing, form filling, and navigating applications. The model achieves 83.0% on the GDPval benchmark^[2] covering 44 cognitive job categories, 75% on the OSWorld test surpassing the human average of 72.4%, and 91% on the legal documents benchmark. Compared to GPT-5.2, error rates dropped by 33%^[3], and the »tool search« mechanism reduces token usage by 47%^[4] in multi-tool environments. The API costs $2.50 per 1 million input tokens^[5] and 10 million output tokens. Simultaneously, Anthropic released computer use functionality within their Cowork and Claude Code packages^[6], also allowing agents to independently operate a computer, while the company cautions that the solution is still early stage and *” Claude may make mistakes^[7], and threats continue to evolve”*. For Polish law firms, financial institutions, and PR agencies, this means reconsidering office-based workflows and the emergence of new AI ops roles replacing traditional execution positions.

Agents Taking Over Computer and Office Processes

With the ability to perform actual actions within corporate systems, the importance of security is rapidly increasing^[8]. On March 18, 2026, Hacker. One launched the Agentic Prompt Injection Testing service^[9], the first commercial tool to test in production whether a prompt injection attack on an agent could lead to real data leaks^[10], code execution, or sending emails. On its platform, Hacker. One recorded a 540% year-over-year increase^[11] in confirmed vulnerabilities of this type, and according to Cisco audits, 73% of production AI deployments contain security gaps^[12]. Another 40% of organizations report having experienced attacks bypassing model protections, and Claw has at least one vulnerability^[13]. Nidhi Aggarwal emphasizes that prompt injection rapidly became a serious threat^[14]. In response, OpenAI published a March 10 guide on designing agents^[15] resistant to such techniques.

Security: Prompt Injection and Deepfake

On March 8, 2026, Bombay Stock Exchange (BSE) issued a second warning^[16] about a fabricated video in which a deepfake impersonated CEO Sundararaman Ramamurthy^[17] promoting a false investment scheme. According to the AI Incident Database analyzed partly by MIT researcher Simon Mylius, deepfake scams have already reached industrial scale^[18]— a China-linked syndicate reportedly extorted^[19] at least 400 crore rupees (approx. 48 million USD) in India, while UK consumer losses reached £9.4 billion^[20] over nine months ending November 2025. One financial firm lost $500,000^[21] in a fake video conference. For Polish PR and corporate communication departments, this represents a real risk of forged executives’ statements, necessitating strict identity verification procedures across digital channels.

EU Regulations and Investment Pressure

At the regulatory level, the AI Act is key. The European Commission published the second draft Code of Practice^[22] for labeling and marking AI-generated content, aiming to become a practical model fulfilling transparency obligations under Article 50 of the regulation. The document promotes a two-tier approach based on secured metadata^[23] and watermarking for text, graphics, audio, and video content. The Code’s finalization is scheduled for June 2026^[24], two months before the AI Act fully takes effect on August 2, 2026. From that date, rules for high-risk systems, market supervision principles, and transparency requirements will apply, with penalties reaching up to €35 million or 7% of global turnover for prohibited practices, and up to 15% for other violations. The European Parliament published an enforcement analysis on March 17^[25], while the platform artificialintelligenceact.eu reminds member states that each must launch^[26] at least one AI regulatory sandbox by August 2, 2026. The implementing act remains underway^[27], including plans to establish the AI Development and Safety Commission as a supervisory body, complicating companies’ preparations for the new obligations.

Capital Concentration and Labor Market Impact

An unprecedented capital concentration in the hands of a few global players continues^[28]. On February 27, 2026, OpenAI announced a $110 billion funding round^[29], in which Amazon invested $50 billion and extended its AWS deal^[30] by another $100 billion over 8 years, with NVIDIA and Soft. Bank each adding $30 billion. The pre-round valuation reached $730 billion^[31], with 2025 revenues at $13 billion and planned expenditures amounting to $115 billion over four years. Meanwhile, ChatGPT now has 900 million weekly users^[32], 50 million subscribers, and 9 million business customers, though the company remains unprofitable and prepares for an IPO. In the same quarter, Anthropic closed a $30 billion Series G round^[33], bringing fresh capital for the two leading frontier model firms to $140 billion combined and significantly raising entry barriers for newcomers.

The labor market consequences of this dynamic are increasingly apparent^[34]. According to analyses, artificial intelligence contributed to a net monthly loss of 5–10 thousand jobs in the US in 2025^[35], with the trend accelerating in 2026. The company Block announced a cut of 4,000 positions^[36] from 10,000 directly citing growing AI capabilities, and Amazon eliminated 16,000 corporate roles^[37]. Data collected by Digital Journal, Reuters, and CNBC indicate^[38] that since early 2026, the global tech sector has lost about 45,363 jobs, with around 9,238 (20%) directly linked to AI deployments and restructuring. Kristalina Georgieva stated in January at Davos^[39] that AI “hits the labor market like a tsunami”, while the share of people fearing job loss due to automation rose^[40] from 28% in 2024 to 40% in 2026. For Polish companies, especially in back-office, customer service, and communication areas, automation is not only technically feasible but also economically justified, despite a lack of national frameworks for systematic workforce retraining.

Related posts:

• AI Accelerates Workforce Restructuring and Regulatory Pressure Ahead of AI Act
• Breakthrough Week in AI: China’s Offensive, Agent Platforms, and Security Crises
• NVIDIA, New Regulations, and a Wave of Funding: The Week AI Agents Enter the Mainstream
• Capital, Agents, and Regulations: A Breakthrough Week in the Global AI Race