Databricks Details Risks From Stolen Personal Access Tokens

AI-enabled attacks are accelerating faster than conventional security measures can respond. CrowdStrike says an agent inside an Argentine healthcare network changed tactics 29 seconds after being blocked and stole five credential hashes 17 minutes later; attackers compromised at least 15 healthcare organizations. The company recorded three times as many agentic intrusions in the second half of 2026 as in the first, and says agents can issue about 1,500 commands an hour. At IRISSCON in Dublin on November 11, Symantec’s Dick O’Brien cited PHP2Shell, exploited 19 minutes after disclosure, and urged organizations to focus on limiting harm—not patching alone. Recommended measures include least privilege, restricting programs, checking outbound data and tying agent actions to named people.

Databricks outlines how a stolen personal access token can lead from an engineer’s account to secrets, service-principal credentials, admin-group membership and exported initialization scripts. It recommends investigating unusual token logins, unfamiliar IPs and unexpected service-principal activity, tracking identities after secrets are accessed, and rotating credentials those identities could reach. Login activity alone may not warrant an alert, and query text is visible only with verbose logging enabled.

Privacy regulators across Spain, Portugal and Latin America say AI vendors generally cannot reuse client personal data to train, evaluate or improve their own models without express permission and a legal basis. They treat this as a separate purpose, say anonymizing data first does not change the analysis, and classify inferred traits such as health or beliefs as sensitive. Separately, HHS awarded about $3.7 million on November 9 to 40 teams developing AI prototypes for Lyme disease and other hard-to-diagnose conditions. The prototypes are not FDA-reviewed devices, and HHS did not say they are cleared for clinical use.

In education, an OpenAI report highlights a gap between student adoption and teacher preparation: a ThinkYoung survey of 5,000 teenagers found 70% use AI for homework, while 40% of teachers have received classroom AI training. Seventy percent of students expect AI to benefit education, compared with 30% of teachers. Parexel chief patient officer Stacy Hurt says about 60% of Americans use tools such as ChatGPT for health benefits and encourages patients to discuss AI-generated information with their care teams. She says effective use could improve research communication, while poor use risks weak communication and study design.

Companies are also testing practical AI tools. At Griffith Foods, Madeline Patterson Smith built a Microsoft Copilot agent for sales teams, limiting it to controlled SharePoint sources and telling it not to guess; the team switched from live Power BI dashboards to month-end snapshots after the agent struggled with current figures. TE Connectivity’s Brendan Stelmach used Copilot to analyze about three dozen files for a corporate responsibility report. The sustainability teams recommend using controlled company sources and checking for overlapping projects. In gaming, Microsoft’s Xbox Labs is testing game reconstruction, browser ports and remixes, while AI tools and modding may make creation easier without ensuring hits. Maurice Heumann’s agent-assisted decompilation work relied on repeated compiling and comparisons; the article also cites planned GTA VI in-game stations, Red Dead Redemption 2 wildlife behavior, and mods inspired by Habbo, Minecraft and Ghost of Tsushima. Counter-Strike began as a Half-Life mod in 1999, but more projects do not guarantee success; Tim Sweeney also described AI-assisted changes to Unreal work.

A November 12 Los Angeles Times letter makes a broader point about accountability: AI systems are programmed tools, and their directors are responsible for their results. The writer recalls a UCLA astronomy team using the SWAC mainframe in the mid-1950s to track two minor planets and two moons of Jupiter, applying “garbage in, garbage out” to today’s AI and its executives.

Key Takeaways

  • CrowdStrike says AI agents can issue about 1,500 commands an hour and recorded three times as many agentic intrusions in the second half of 2026 as in the first.
  • Attackers compromised at least 15 Argentine healthcare organizations; one AI agent changed tactics 29 seconds after a blocked attempt and stole five credential hashes 17 minutes later.
  • Symantec’s Dick O’Brien cited PHP2Shell exploitation 19 minutes after disclosure and urged defenses focused on limiting harm, least privilege and accountability for agent actions.
  • Databricks warns that a leaked personal access token can expose secrets, enable privilege escalation and lead to data exports; it recommends tracking affected identities and rotating accessible credentials.
  • Regulators in Spain, Portugal and Latin America say vendors generally need express permission and a legal basis to reuse client personal data for their own AI models.
  • HHS awarded about $3.7 million to 40 teams building AI prototypes for Lyme disease and other hard-to-diagnose conditions; they are not stated to be cleared for clinical use.
  • An OpenAI education report cites a survey in which 70% of students use AI for homework, while 40% of teachers have received training on classroom use.
  • Parexel’s Stacy Hurt says about 60% of Americans use AI tools such as ChatGPT for health benefits and recommends discussing AI-generated information with care teams.
  • Griffith Foods and TE Connectivity used Microsoft Copilot for sustainability work, with teams emphasizing controlled sources, verification and awareness of duplicate projects.
  • Microsoft’s Xbox Labs is testing game reconstruction, browser ports and remixes; AI and modding may ease creation, but the article says they do not predict which games will become hits.

AI Agents Move Through Networks Faster Than Old Security Measures

CrowdStrike reported that an AI agent inside an Argentine healthcare network changed tactics 29 seconds after its first attempt was blocked, then stole five credential hashes from a domain controller 17 minutes later. The attacker compromised at least 15 healthcare organizations in Argentina and took credentials and medical records from at least one. CrowdStrike says AI agents can issue about 1,500 commands an hour, compared with dozens from human attackers, and its threat hunters recorded three times as many agentic intrusions in the second half of 2026 as in the first half. The speed has led CrowdStrike to reconsider “breakout time,” a measure of how long attackers take to move between network devices, but the company has not agreed on a replacement. Other possible measures track attacker command speed and handoffs, as well as defenders’ detection and response times.

Security Leaders Urge Mitigation as AI Shrinks the Time to Patch

At the IRISSCON 2026 conference in Dublin on November 11, Symantec analyst Dick O’Brien said organizations cannot rely on patching alone as attackers use AI. The PHP2Shell vulnerability was exploited just 19 minutes after it became public in October 2026. O’Brien expects attackers to move toward fully agent-run attacks over the next year, making threats harder to detect and stop. He advised organizations to assume vulnerabilities will be exploited and measure how quickly they can reduce harm, rather than how quickly they can install patches. His suggested steps include blocking harmful behavior, limiting what programs can run, using least privilege for agents and tokens, checking data before it leaves devices, and linking each agent action to a named person.

HHS Awards $3.7 Million to AI Prototypes for Lyme and Other Illnesses

The Department of Health and Human Services awarded about $3.7 million on November 9 to 40 teams developing AI prototypes for Lyme disease and other hard-to-diagnose conditions. The TOPx HHS Tech Sprint for AI and Invisible Illness was led by the HHS LymeX Innovation Accelerator, the National Institutes of Health Office of Research on Women's Health, and the Federal Chief Data Officers Council. The competition received 171 submissions, with 46 teams advancing to build and demonstrate prototypes. Prizes included two $1 million grand prizes, five $200,000 awards, six $50,000 awards, and 27 impact awards of $15,000 each; another $50,000 in People's Choice awards will be announced later. These are prototypes, not FDA-reviewed medical devices, and HHS did not say they are cleared for clinical use.

AI Helps Game Makers Build More, but Hit Games Remain Uncertain

A StartupHub.ai article discusses how AI tools and game modding may make it easier to create games, while offering no clear sign that more projects will become hits. It describes Maurice Heumann's first-person shooter decompilation work, which used agents and repeated compiling and comparisons to preserve how the game behaved. The article also points to planned in-game stations for GTA VI, wildlife behavior in Red Dead Redemption 2, and mods inspired by Habbo, Minecraft, and Ghost of Tsushima. Counter-Strike began as a Half-Life mod in 1999, showing how mods can lead to major games, but the article says an abundance of projects does not predict success. Microsoft has formed Xbox Labs to test game reconstruction, browser ports, and remixes, while Tim Sweeney described changes to Unreal work that use AI assistance.

Letter Says AI Outcomes Reflect the People Who Program It

In a letter published by the Los Angeles Times on November 12, 2026, a reader argued that AI systems are not rogue actors and that their directors are responsible for their results. The writer recalled working in the mid-1950s for a UCLA astronomy professor, whose team used the SWAC mainframe to track two minor planets and two moons of Jupiter. The postdoctoral researcher who ran the office coded the computer to provide the needed information. The letter applies the old phrase “garbage in, garbage out” to AI, saying that technology executives who direct its programming and use are responsible for its output.

Ibero-American regulators say client data cannot train vendor AI by default

Data protection regulators across Spain, Portugal and Latin America published AI privacy guidance in November 2026. It says an AI vendor or data processor generally cannot use personal data handled for a client to develop, train, evaluate or improve its own models unless it has express permission and a legal basis. The guidance says that use counts as a different purpose, and anonymizing the data first does not change the analysis. It also says inferred traits such as health or beliefs count as sensitive data, and simply telling people that AI is used does not count as proper notice. The standards complement existing data protection rules and do not replace national laws.

Databricks details how stolen credentials can hide in normal activity

Databricks described how attackers can use a leaked personal access token to move from an engineer’s account to other identities and data. In its example, a token taken from a misconfigured CI/CD log is used to access secrets, including a service principal credential, which is then used to join an admin group and export initialization scripts. The authors recommend looking for unusual token logins, new IP addresses, unexpected service principal activity and bursts of discovery API calls. They say teams should follow both identities after a secret is read and rotate credentials that the suspect identity could access. Databricks also warns that login activity alone rarely justifies an alert, and query text is only available when verbose logging is enabled.

Parexel leader says patient AI use should shape clinical research communication

Stacy Hurt, chief patient officer at Parexel, told Pharmaceutical Executive that AI could help improve communication between clinical research sponsors and patients. Hurt said about 60% of Americans use AI tools such as ChatGPT for health benefits. She encourages patients to discuss AI-generated information about their diagnosis and care plan with their healthcare team. Hurt said well-used AI can help research teams better understand patient views and support deeper conversations, but poor use could lead to badly designed studies or weak communication.

OpenAI report highlights teacher training gap as students use AI for homework

OpenAI’s Responsible AI in Education report points to a gap between student use of AI and teacher guidance. The article reports that ThinkYoung surveyed 5,000 teenagers, with figures showing 70% of students use AI for homework while only 40% of teachers have received training on using it in class. The survey also found that 70% of students expect AI to have a positive effect on education, compared with 30% of teachers. The report calls for educators, policymakers and industry leaders to provide more teacher training and support.

Sustainability teams share practical lessons from building AI tools

Trellis Network members shared lessons from creating AI tools for sustainability work, often using software their companies already license, such as Microsoft Copilot, Google NotebookLM and OpenAI ChatGPT. At Griffith Foods, Madeline Patterson Smith built a Copilot agent to help sales teams discuss sustainability and nutrition, restricting it to controlled SharePoint sources and instructing it not to guess. The team replaced live Power BI dashboards with month-end snapshots after the agent struggled to identify current figures, and found that it favored information stored in Word documents. At TE Connectivity, Brendan Stelmach used Copilot to analyze about three dozen files in different formats while preparing for the company’s next corporate responsibility report. The article recommends limiting agents to company-controlled sources and checking whether another team is already working on a similar tool.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

AI cybersecurity agentic AI attacks healthcare cybersecurity credential theft least privilege incident response personal access tokens AI data privacy AI model training healthcare AI AI education Microsoft Copilot enterprise AI AI in gaming game modding AI accountability AI governance sustainability reporting

Comments

Loading...