Lakera Bulletin - This Week in AI: When AI Cyber Risk Goes Systemic

[Askal]

It’s been a week of big model releases and bigger security questions: from OpenAI’s new flagship and Anthropic’s latest update to mounting warnings from governments, regulators, and banks. Across the board, the theme is clear: capabilities are advancing fast, and the pressure to secure them is rising just as quickly.

Let’s get into it.

OpenAI Releases GPT-5.4

OpenAI unveiled GPT-5.4 as its most capable and efficient frontier model for professional work, bringing together stronger reasoning, coding, and agentic workflows in a single system. The release also adds native computer-use capabilities, better tool use, and stronger performance across spreadsheets, presentations, documents, and deep web research.
🔗 Read the official announcement

Anthropic Releases Claude Opus 4.7

Anthropic’s newest flagship model improves performance across coding, reasoning, and multimodal tasks, while introducing stronger safeguards against high-risk misuse. The release signals continued progress toward more capable,  and more tightly controlled, frontier systems.
🔗 Read the official announcement

U.S. Government Considers Deploying Mythos

The White House is evaluating controlled access to Anthropic’s Mythos model across federal agencies for vulnerability detection. The move reflects growing urgency to adopt AI defensively, even as concerns about dual-use risks remain unresolved.
🔗 Read the full story

UK Warns of AI-Driven Cyber Threat Surge

UK officials issued an open letter warning that AI can now discover and exploit vulnerabilities at unprecedented speed. Businesses are being urged to treat AI cyber risk as a board-level priority.
🔗 Read the letter

Bank of England Stress-Tests AI Risk

The Bank of England is running simulations to understand how AI agents could destabilize financial systems or amplify cyber threats. Regulators warn that risks could scale rapidly as adoption accelerates.
🔗 Read the report

Goldman Sachs Flags AI Cyber Risk

Goldman Sachs leadership warned that advanced AI models could expose vulnerabilities across shared financial infrastructure. The concern highlights how cyber risk is becoming a systemic issue across the global economy.
🔗 Read the coverage

ICYMI: AI Red Teaming for Non-Deterministic Systems

AI breaks traditional security assumptions: systems are probabilistic, and vulnerabilities can emerge through subtle changes in prompts, context, or model updates. Lakera’s latest post explores why red teaming must become continuous, application-specific, and focused on real-world agent behavior rather than static tests.
🔗 Read the blog

The Enterprise Playbook for Agentic AI Security

AI systems now retrieve data, invoke tools, and act across enterprise workflows.
Get the playbook to learn how to secure AI across employees, applications, and agents.

👉 Explore the Playbook

From new defensive models to growing fears of systemic risk, this week makes one thing clear: AI security is no longer a niche concern — it’s becoming foundational to how institutions operate.

See you next week!