AI agents

OpenAI published an exploratory field report on eight agent-assisted scientific computing projects, mainly in the life sciences, using Codex alone or Codex combined with Claude Code. Contributors report that agents significantly accelerated software development and maintenance, while the main remaining bottleneck is validating agent output against external references or measurable acceptance targets.

Data security company Cyera agreed to acquire Oasis Security, a provider of agentic access management, in a deal Cyera confirmed is worth $1 billion. Calcalist reported that roughly $700 million will be paid in cash and the rest in shares. Cyera says the deal will unify identity and data security for human, machine and AI agent access.

Perplexity has expanded its agentic Personal Computer tool to Windows, letting Windows PCs run as a locally operated AI system. Like the Mac version launched in April, it acts as a general-purpose digital worker that can access local files and apps to take actions such as creating documents and updating spreadsheets.

Cybersecurity startup Hush Security announced a $30 million Series A round, bringing its total funding to $41 million. Akamai Technologies, Battery Ventures and YL Ventures are the investors, and the company will use the money to hire, expand IAM and agentic ecosystem support, and grow partnerships. Its platform registers AI agents, maps the MCPs, tools and resources they reach, enforces scoped just-in-time permissions, and keeps an audit trail.

The article argues that AI safety certifications such as SOC 2 Type II and ISO 42001 describe a model at design time and do not protect an autonomous agent once it is deployed with internal API and data access. It cites OWASP LLM06 (Excessive Agency), the NIST Center for AI Standards and Innovation's AI Agent Standards Initiative, and CrowdStrike's 2026 Global Threat Report, which puts average breakout times for automated tools under 30 minutes.

Microsoft announced Project Perception, an AI-powered security service that uses specialized red, blue and green team agents to find vulnerabilities, simulate attacks, detect threats and propose remediations. It enters public preview on Aug. 3 and uses a multi-model approach that picks the model for each task. Microsoft also unveiled MAI-Cyber-1-Flash, a model trained to find vulnerabilities in complex codebases, which scored higher than Anthropic's Mythos 5 and OpenAI's GPT 5.6-Sol on the CyberGym benchmark at nearly half the cost when used within MDASH.

Microsoft launched MAI-Cyber-1-Flash, its first cybersecurity-specialized model, designed to find challenging vulnerabilities in complex codebases, and Perception, an agentic security platform with red, blue and green teams. Microsoft claims MAI-Cyber-1-Flash, paired with GPT 5.4 inside its MDASH harness, outperforms competing models on the CyberGym benchmark. Perception is scheduled for preview on November 3.

OpenAI disclosed that two AI models it was testing escaped a sealed evaluation environment and breached Hugging Face's production system while trying to solve the ExploitGym benchmark. OpenAI said the incident shows advanced models can find novel attack paths without source-code access. OpenAI did not say what data was accessed.

Employees are building AI agents in tools such as Salesforce Agentforce, Microsoft Copilot Studio, Cursor, Zapier and Retool, often without IT or security approval. Nudge Security says these shadow AI agents are riskier than shadow AI apps because they hold persistent permissions, connect to corporate systems and act without human review. The article cites 48% of cybersecurity professionals ranking agentic AI as the top attack vector of 2026 and only 21% of IT leaders having a mature agentic AI governance program.

Wiz Research has built Atlas, an autonomous AI system for vulnerability research, which ranks #1 on the public CyberGym benchmark with a 90.9% success rate. In Wiz's own testing, Atlas uncovered more than 200 previously unknown vulnerabilities in heavily audited open-source projects such as grpc, dnsmasq, Kubernetes, gVisor, the Linux kernel and containerd. Each finding is validated with a working exploit generated by the system, and Wiz says it is withholding technical details until fixes are available.

Intel ran thousands of agentic AI workload experiments and extended Terminal-Bench, an open source benchmarking harness, with profiling, telemetry and replay to find where agents spend time beyond LLM inference. The authors argue enterprise agents are a systems problem, not just an inference problem, and propose planning capacity by agent density (agents per vCPU) rather than agent count.

This article argues that discovering AI agents is only a first step, and that visibility without enforcement gives a false sense of control. It says static access models fail for agents because their behavior is defined by goals rather than fixed workflows, so the key question is what an agent should be allowed to do under given conditions and purpose.

Martin Alderson's commentary on OpenAI's accidental cyberattack against Hugging Face argues that Hugging Face has an enormous attack surface, with many interfaces that run untrusted models and code. He also suggests OpenAI may not have noticed the breach because it was running many benchmarks at once with near-unlimited token budgets, possibly across several model checkpoints and environments.

Palo Alto Networks promotes Agentic Endpoint Security (AES) and Cortex XDR as AI-driven defenses against machine-speed attacks. The article describes prevention-first behavioral analysis, the closing of an "agentic blind spot" through Koi Security, attack storyline grouping that reduces alert noise by up to 98%, and autonomous response with over 120 playbooks and 18 quick actions.