Goldman Collaborates with Anthropic to Address AI Cyber Risks

Goldman Sachs is working directly with Anthropic following concerns about the capabilities of Anthropic's new model, Mythos. David Solomon, Goldman Sachs CEO, said the bank is doubling down on cyber protections as the model's advanced abilities raised alarm among officials and experts.

The collaboration targets model-driven attack surfaces that matter to financial institutions. Key risk classes include automated social engineering, high-quality phishing content generation, code and exploit synthesis, and tools that accelerate vulnerability discovery. Practical defensive levers likely under evaluation include:

• •robust red-teaming and adversarial testing against Mythos-style capabilities
• •API access controls, rate limits, and fine-grained usage policies
• •content filtering, prompt engineering constraints, and response sanitization
• •model watermarking and forensic telemetry for provenance and output detection
• •deployment segregation such as private-hosted models or vetted on-prem instances

This is a material shift in how large financial institutions engage with frontier-model vendors. Banks have long been high-value targets for adversaries using automation. The arrival of models with improved reasoning and code synthesis increases the attack automation multiplier. Institutional responses will shape vendor product requirements: hardened APIs, contractual security SLAs, expanded red-team programs, and stronger detection tooling. For the AI vendor ecosystem, enterprise adoption will increasingly depend on demonstrable safety engineering, transparent risk assessments, and joint incident playbooks.

Monitor whether collaborations produce standardized safety checklists or new contractual controls for enterprise model access, and whether regulators ask for formal risk assessments for models used in critical infrastructure.

Bottom line: The Goldman-Anthropic engagement signals that threat modeling and operational safeguards are now a procurement requirement for frontier models in finance. Security teams should prioritize adversarial testing, telemetry, and access governance when evaluating large generative models.