AI Sec Watch

Deepfake voice and video attacks (AI-generated replicas of real people) are becoming increasingly common and costly, with tools that require only three seconds of audio and cost almost nothing to create. Attackers target finance employees and IT staff by impersonating executives on calls or video meetings to authorize large money transfers or credential changes, and these attacks bypass traditional security tools because they rely on tricking people rather than exploiting software vulnerabilities. Organizations that have successfully stopped these attacks all used the same defense: training employees to pause and verify requests before acting on them.

Some people worry that advanced frontier LLMs (large language models, AI systems trained on massive amounts of text) like Claude Mythos and GPT-5.5 could cause serious cybersecurity problems by being misused for attacks. However, security researcher Ari Herbert-Voss suggests this situation could also present opportunities.

Microsoft's 'Agent ID Administrator' role, designed to let AI agents have controlled identities in Entra ID (Microsoft's identity management system), had a security flaw that let users take ownership of unrelated service principals (the tenant-specific identities that applications use to authenticate and access resources). This meant attackers could gain the same privileges as more powerful administrator roles and potentially take over the entire tenant (organization's cloud environment).

DeepSeek released V4, a new AI model that can process longer text more efficiently and matches the performance of leading competitors from OpenAI, Anthropic, and Google while remaining open source. Researchers are increasingly focused on developing world models (AI systems that understand and can interact with the physical world, not just digital tasks) to overcome limitations of current language models and enable advances in robotics and physical tasks like laundry folding or navigation.

Google researchers found that indirect prompt injection attacks (hidden traps where malicious instructions in external data trick AI systems into bypassing their safety rules) on websites are increasing, with a 32% rise between November 2025 and February 2026, but current attacks remain relatively unsophisticated. The attacks they discovered fell into two categories: exfiltration attempts that try to steal data like IP addresses and credentials, and destruction attempts that aim to delete files, though neither showed advanced techniques. Researchers warn that while today's attacks are low in sophistication, the upward trend suggests the threat will soon grow in both scale and complexity.

Anthropic's Claude Mythos is an AI system that can discover vulnerabilities much faster than human teams, but organizations are unprepared for the remediation (fixing) side of the process. The real problem isn't finding vulnerabilities quickly, it's that most teams lack the infrastructure to triage, prioritize, and verify fixes once they're discovered, so faster discovery just creates a growing backlog of unfixed critical issues.

AI is transforming DevSecOps (the practice of integrating security into software development processes) by embedding security checks earlier in coding and automating vulnerability detection and fixes. The shift moves security from happening after code is written to happening during code generation itself, with AI tools providing secure coding guidance, scanning for vulnerabilities using reasoning rather than fixed rules, and suggesting automated fixes integrated directly into developer workflows.

Rather than eliminating SOC analyst jobs, agentic AI (AI systems that can independently execute tasks) is transforming entry-level analysts from performing repetitive investigative work into 'managers of agents' who oversee AI systems and make decisions based on their findings. The shift moves analysts from manually gathering evidence across multiple systems to reviewing AI-generated investigations and validating conclusions, allowing them to handle more alerts at a higher level of judgment.

Elon Musk is suing Sam Altman and OpenAI, claiming they violated their founding agreement by converting OpenAI from a non-profit (an organization that doesn't aim to make money for owners) to a for-profit business. The lawsuit alleges fraud and breach of contract, with the trial beginning in Oakland, California, and expected to last two to three weeks.