Google's Gemini Hacked Three Companies in New AI Safety Incident

Google's Gemini was used to compromise three companies in a newly documented AI safety incident, according to the Financial Times, marking one of the most concrete examples yet of a frontier language model being weaponised against real enterprise targets. The incident adds urgency to ongoing debates about agentic AI safety at a moment when Anthropic's financial trajectory is drawing intense scrutiny from pre-IPO investors.

What Happened

The Financial Times reports that attackers exploited Google's Gemini model to successfully breach three separate companies, in what the publication describes as a new AI safety incident. The report does not name the affected organisations but indicates the intrusions were enabled by Gemini's reasoning and agentic capabilities, underscoring the dual-use risk inherent in increasingly capable multimodal and reasoning-focused models. The incident is likely to intensify pressure on Google and other frontier AI developers to implement tighter controls around autonomous task execution, particularly as models are increasingly embedded in enterprise workflows with access to sensitive systems. The story connects directly to the broader challenge of AI agents in the enterprise operating with limited human oversight.

In a separate but related finding, the Financial Times published research showing that AI chatbots give wrong answers to financial queries most of the time. The study tested leading large language models on common personal finance and investment questions and found accuracy rates that fell well below what users might reasonably expect. The finding is significant because financial advice represents one of the highest-stakes domains for LLM deployment, and widespread public use of these tools without appropriate guardrails carries material risk. The chatbots tested were not identified by name in available summaries, but the study's framing implicates the category broadly, including GPT-class and Gemini-class systems.

Financial Times reporting on Anthropic's financial position adds further context: investors are weighing whether the company can sustain its surging revenues after a potential IPO. Anthropic has reported rapid revenue growth driven by its Claude model family, but questions persist about margin sustainability given the compute intensity of frontier model inference. The publication notes that post-IPO scrutiny will focus on whether enterprise adoption can keep pace with the company's current valuation trajectory. The company did not disclose specific revenue figures in the reporting available.

On the technical side, MarkTechPost published a detailed explainer on the four dominant LLM quantisation and packaging formats in active use: GGUF, GPTQ, AWQ, and EXL2. The piece provides a timely reference as model deployment complexity grows, with each format offering distinct tradeoffs between inference speed, memory footprint, and compatibility across hardware backends. GGUF has emerged as the standard for CPU-offloaded inference, while AWQ and EXL2 are preferred for GPU-accelerated deployments where throughput per token matters most. The format field reflects the ongoing engineering challenge of making frontier-scale reasoning models practical at the edge and in cost-constrained inference environments.

Why It Matters

The Gemini hacking incident is not an abstract red-team exercise. Three real organisations were compromised, and the vector was a production-grade reasoning model with agentic capabilities. This signals that the safety gap in multimodal and reasoning-focused AI is not theoretical: as models gain the ability to plan, execute multi-step tasks, and interact with external systems, their attack surface expands in kind. The financial chatbot accuracy findings compound this picture. If frontier LLMs fail on structured, verifiable financial queries most of the time, then their deployment in high-stakes advisory contexts is outpacing what the technology can reliably support. Investors tracking AI infrastructure investment dynamics will note that the gap between capability claims and demonstrated reliability is becoming a material risk, not just a technical footnote. Anthropic's IPO calculus sits squarely in this environment.

Background

The reliability and safety of large language models have been active research areas since GPT-4's release introduced multimodal reasoning at scale. The progression toward agentic systems, where models autonomously plan and execute sequences of actions, has accelerated since 2024. For a technical grounding in how multimodal reasoning architectures are constructed, the MachineHours encyclopedia entry on GPT-4 Multimodal provides useful context on the architectural decisions that underpin current frontier systems.

The quantisation format debate tracked by MarkTechPost reflects a parallel strand of AI research focused on inference efficiency. As reasoning models grow larger, the ability to compress and deploy them without catastrophic accuracy loss has become a critical area of applied research, directly affecting which organisations can afford to run frontier inference in production.

What to Watch

Google has not yet disclosed how the Gemini-enabled breaches occurred at a technical level, and the company's response, including any model updates or access restrictions, will be closely watched by enterprise security teams. Regulatory bodies in the EU and US are likely to cite this incident in upcoming AI liability discussions. For Anthropic, the path to a sustainable post-IPO revenue base depends on enterprise trust, which incidents like the Gemini breach, even if involving a competitor, will complicate across the sector. The accuracy findings on financial chatbots may also accelerate calls for mandatory disclosure requirements when LLMs are used in regulated advisory contexts. Watch for updated safety benchmarks from both Google DeepMind and Anthropic in Q4 2026.