Autonomous AI agents dominated headlines this week as fast-moving security exploits and new agent-capable tools surfaced — from shopping and transaction management to scientific research coordination. On the latest episode of This Week in AI, host Vicki Reyzelman of Akamai explored how these trends are stressing existing security practices and opening business and research opportunities.
Attacks are outpacing traditional defenses
AI is changing how attackers discover and exploit vulnerabilities. A supply-chain compromise that leveraged ARTEX, an open-source agentic AI penetration-testing framework, reached more than 2,250 Microsoft 365 tenants via a ConnectWise compromise. That incident highlighted limitations in OAuth authentication and conventional identity protections when autonomous agents interact with enterprise systems.
A more immediate high-severity issue came from PHP2Shell: two vulnerabilities rated 9.2 on the CVSS severity scale that could allow unauthenticated code execution on vulnerable web servers. The Cybersecurity and Infrastructure Security Agency (CISA) urged immediate patching following the October 21 disclosure. Researchers also documented attackers using AI agents to obtain password hashes and multi-factor authentication secrets from a European government system within two hours.
The speed of exploitation leaves organizations little time to investigate and respond, especially smaller businesses with limited security resources. Security teams can no longer rely solely on manual review: they need automated detection and response, tighter access controls, and network segmentation to limit what compromised agents can reach. Monitoring autonomous agents, however, is itself difficult.
Researchers from METR and Redwood Research identified weaknesses in an earlier version of OpenAI’s internal monitoring system, including a flaw that could allow hundreds of agent actions to be concealed. OpenAI says it has patched vulnerabilities affecting production safeguards, though the researchers have not independently evaluated those fixes.
The AI-safety debate is expanding beyond purely technical measures. Anthropic proposed federal AI safety legislation, and Google DeepMind employees resigned in protest over the company’s military AI work. Both developments reflect growing pressure on AI companies to clarify rules for how their technology is developed and used.
The shift to agents creates new business and oversight challenges
Retailers are taking different approaches to autonomous shopping agents. Walmart introduced an Agent Gateway that allows approved agents to shop within customer-defined spending limits and approval requirements. Amazon has restricted some third-party shopping agents, while banks including Morgan Stanley and Bank of America are limiting agent access to customer accounts.
The disagreement goes beyond security: retailers that rely on advertising and product recommendations have incentives to keep shoppers on their own sites. As more purchases start with agents, companies will need to rethink how they maintain customer relationships.
Anthropic is experimenting with another oversight model. Claude’s auto mode is becoming the default for Pro, Max, and Team users, enabling automated checks that approve routine actions and block risky ones. Anthropic also contributed its Agent Skills format to the Agentic AI Foundation under the Linux Foundation; more than 70 platforms have adopted the standard and over a million skills have been published. Shared formats could make workflows easier to reuse across products, but organizations still need to control which actions those workflows can perform.
Infrastructure and energy constraints matter
Growing agent use increases compute and electricity demand. NVIDIA has announced $164.5 billion in commitments and guarantees to support AI data centers, including an Ohio campus that SB Energy is building for OpenAI. Power shortages, however, have delayed other projects: a proposed data center in Essex was stalled because grid operators could not promise adequate electricity until the 2030s.
Meanwhile, Amazon is backing a 960-megawatt nuclear project, and Brookfield and NextEra are pursuing a $100 billion AI hub at a former nuclear site in Kentucky. But money and chips alone cannot solve AI infrastructure constraints when expanding generation and grid capacity can take years.
Faster discovery, but validation still takes time
Healthcare offered an encouraging example of AI’s potential. A study of Microsoft’s MAI Diagnostic Orchestrator found it correctly diagnosed 83.2% of difficult medical cases, compared with 36.1% for generalist physicians. The comparison did not include specialists, and the system still needs clinical validation, but the research suggests AI could help clinicians investigate challenging cases and identify conditions they might otherwise miss.
Drug discovery could also benefit from more capable agents. OpenAI announced its acquisition of Orchestra Bio, a platform that coordinates drug-development workflows, with plans to integrate it with Rosalind, its life-sciences research tools. That combination could help researchers manage experiments and collaborate more efficiently within a development process that typically takes 10 to 15 years from initial discovery to approval.
An Anthropic protein-design experiment offered a concrete example of accelerated early-stage research: in 64 hours an Anthropic model generated 318,484 candidate protein designs targeting scorpion venom. Researchers from Adaptyv Bio and Twist Bioscience subsequently tested 211 designs and identified 17 that bound to the toxins. The researchers have not established whether those proteins neutralize venom, but the experiment demonstrated how AI can rapidly narrow an enormous search space before more costly laboratory testing begins.
What’s next
This Week in AI will return Monday, November 2, with another episode covering the latest developments in AI research and applications. Previous episodes are available on YouTube, Spotify, Apple, and other podcast platforms.
