Friday, 20 November 2026 Follow @AIDailyFive

The AI Daily Five · No. 19 · Wednesday 4 November

Who Decides When AI Is Safe?

Today’s edition examines the compromised Claude Code extension in Cursor, restrictions on Google DeepMind researchers, Google’s AI controls for schools, CrowdStrike’s agent security tools and a proposed clinical laboratory system.

Two young people with tape over their mouths at laptops, with students coding and a scientist at a microscope behind them.
Illustration: AI Story Front design team

1

Fake Claude Code Extension Hits Cursor. Who Can Developers Trust?

A malicious extension reached developers through Cursor’s marketplace, carrying software designed to establish persistent access to their computers. Subsequent technical analysis shows how an apparently legitimate update could expose sensitive development environments.

For about 11 hours on 29 October, attackers distributed malicious versions of Anthropic’s Claude Code extension through the Cursor VS Code Extension Marketplace. The packages were designed to install an SSH backdoor, potentially giving attackers remote access to developers’ computers.

According to Cursor’s security notice, the affected versions were 2.1.312 and 2.1.313 for macOS and Windows. The company removed them, temporarily shut down its extension marketplace and worked with the hosting provider to disable the attacker’s server.

This was a software supply-chain attack. Anthropic told Microsoft that its systems were not compromised and that it did not develop the malicious package. Microsoft says its Visual Studio Marketplace and the official Claude Code npm package were not affected.

On 30 October, Microsoft Security Research published a technical investigation explaining how the malicious code operated.

The extension attempted to collect information about developers’ computers, add an attacker-controlled SSH key and establish a reverse connection to a remote server. It also created mechanisms intended to restore that connection after a computer restarted.

Cursor says errors in the code and administrator-permission requirements blocked SSH access on most macOS and Windows machines. However, the malware could still leave unauthorised keys and startup scripts behind, meaning that uninstalling the extension alone might not remove the threat.

BleepingComputer reported that up to 5,000 developers may have received an affected version. That figure describes potential exposure, not confirmed successful compromises of 5,000 computers.

Cursor has published remediation instructions for affected users. The company says the attacker’s server has been disabled and that its investigation found no evidence of other tampered extensions.

The incident demonstrates a particular vulnerability in AI-assisted software development. Coding assistants operate inside environments containing source code, credentials and access to corporate systems. A compromised extension can exploit those permissions without needing to defeat the AI model itself.

2

Can AI Researchers Speak Freely About Risks?

Google DeepMind employees are challenging contractual restrictions that they fear could discourage public criticism long after they leave. The dispute exposes a tension between corporate confidentiality and the public interest in AI safety.

Employees at Google DeepMind in the UK are challenging employment clauses that they believe could prevent them from criticising the company after leaving their jobs, according to a Washington Post investigation published on 3 November.

According to a contract reviewed by the Post and two current workers, some London staff are barred from criticising the company for life, even after leaving. Some researchers worry the provisions could discourage discussion of AI’s risks, even when they believe disclosure would serve the public interest.

The Washington Post investigation reports that dozens of employees recently joined an internal discussion about the clause, according to one current employee.

The concerns intensified after former DeepMind employee Sofia Liguori publicly criticised Google’s involvement in military AI despite contractual language she believed might restrict such comments.

Google says it does not enforce post-employment restrictions on criticism and has removed the relevant provision from contracts for new employees. However, some existing staff want the clause dropped from their contracts too, rather than relying on assurances about enforcement.

The dispute comes amid broader disagreements within AI companies about how research risks should be communicated.

Researchers working on advanced AI may have access to information about model capabilities, limitations and potentially harmful behaviour that is not available to the public.

Companies have legitimate interests in protecting confidential research, security information and intellectual property. But broad restrictions on criticism can create uncertainty about whether employees may discuss matters affecting public safety.

The distinction between disclosing confidential technical material and expressing concerns about a company’s decisions is particularly important.

Google maintains that employees have internal channels for raising concerns. The disagreement concerns whether those mechanisms provide sufficient protection when workers believe public discussion is necessary.

3

Can Schools Refuse Google’s AI?

Google is giving school administrators new controls over AI services, but conventional Search may become unavailable to pupils if those features are disabled. The policy raises questions about how much choice schools really have.

Google is introducing new controls for its education products that could leave pupils without access to Google Search if their schools decline to enable AI-integrated services.

Google faced an outcry in August after switching on Gemini for many K-12 students.

Announced on 3 November, the changes affect Google Search, Gemini, Gemini Notebook and Gemini in Classroom for school-managed accounts belonging to pupils in kindergarten through Year 12 or equivalent.

Under Google’s new education policy, these services will initially be switched off until school administrators enable them. Google says it will share specifics with schools in the coming months, ahead of the rollout. The company says the approach responds to concerns from educators and gives schools greater control over AI use.

However, Google Search itself now incorporates AI features. Schools that do not enable the relevant services risk losing Search access through those student accounts, rather than retaining a version of Google Search without generative AI.

The distinction has attracted criticism from educators and policymakers. As the Associated Press reported, critics argue that schools should not have to choose between a widely used search engine and AI functionality they may not consider appropriate for their pupils.

Google says administrators will have individual controls for its products. The company is also introducing Guided Learning capabilities intended to support pupils through explanations and questions rather than immediately supplying answers.

Guided Learning is now available globally through Google Classroom, while a pilot integrating it into Search is planned for selected US schools later this month.

The company argues that these tools can support understanding and reduce reliance on AI-generated answers. Its claims about educational benefits should be distinguished from independent evidence of improved learning outcomes.

Google Classroom is used by schools worldwide, giving decisions about its features considerable influence over how educational technology is introduced.

The new controls are intended to provide schools with more authority over AI access. Yet the integration of Search and AI complicates their ability to reject one without losing the other.

4

Can CrowdStrike Stop Rogue AI Agents?

The cybersecurity company has announced tools to interrupt unauthorised agent activity and simulate attacks before criminals exploit vulnerabilities. Both developments reflect the difficulty of securing software that can act independently.

CrowdStrike has introduced new security capabilities designed to monitor autonomous AI agents, interrupt potentially harmful actions and simulate cyberattacks against corporate systems.

The announcements, made on 3 and 4 November during its Fal.Con Europe conference, reflect a growing concern among cybersecurity companies: AI agents are gaining access to sensitive information and business systems, often with permission to perform tasks without constant human supervision.

On 3 November, CrowdStrike announced an AI Kill Switch within Falcon Guardian, designed to stop unauthorised agent activity while it is running. The Kill Switch works first at the AI gateway, with endpoint support to follow.

The company says the system combines monitoring at the agent level with network inspection and controls on the computers where agents operate. Rather than relying entirely on instructions embedded in an AI model, it aims to interrupt activity through the surrounding security infrastructure.

CrowdStrike also announced plans to acquire Fiddler AI, a company specialising in monitoring AI behaviour, and a definitive agreement to acquire Dreadnode, which develops automated security testing for AI models and agents. These are planned acquisitions, not completed integrations.

A day later, the company introduced Agentic Adversary Emulation, a system designed to imitate how attackers exploit multiple weaknesses to reach valuable systems. The tool is available only in early access for select customers.

Powered by its SafeMind technology, the tool is intended to identify possible attack paths, recommend which vulnerabilities to address first and repeat simulations to assess whether the weaknesses have been closed.

CrowdStrike contrasts this with point-in-time penetration tests and periodic assessments, which it says cannot keep pace.

CrowdStrike presents the technologies as complementary: one attempts to stop dangerous behaviour during operation, while the other seeks weaknesses before attackers exploit them.

However, the announcements do not establish how reliably the systems perform across independent real-world evaluations. Their effectiveness will require evidence beyond the company’s own demonstrations.

As businesses deploy agents capable of accessing software, managing information and executing tasks, security controls increasingly need to monitor actions rather than merely assess the underlying model.

5

Can AI Prevent Clinical Laboratory Errors?

Researchers propose coordinating specialised AI agents to monitor laboratory equipment, sample quality and testing delays. Their framework could help identify problems earlier, but it has not yet been validated in clinical practice.

Researchers at Weill Cornell Medicine have proposed an AI framework designed to identify problems in clinical laboratories before they disrupt testing or affect patient care.

The system, called Laboratory Agentic Reasoning and Integrative Analytics (LabARIA), would coordinate specialised AI agents across different laboratory operations. Unlike tools designed to analyse individual test results, it would monitor the processes involved in producing those results.

The researchers describe the approach in a paper published in npj Health Systems on 28 October. Weill Cornell announced the work on 2 November.

Clinical laboratories analyse blood, tissue and other samples used to diagnose disease and monitor treatment. Their reliability depends on equipment performance, sample handling, staffing and quality-control procedures.

Existing software can monitor individual parts of these operations. However, the researchers argue that information often remains fragmented between laboratory systems, making it difficult to identify problems developing across several processes.

LabARIA would introduce a coordinating system capable of gathering information from existing laboratory databases and directing it to specialised agents.

Individual components could monitor instrument performance, identify changes in sample quality, track testing delays and anticipate shortages of supplies or staff. Supervisory agents would combine the findings and alert laboratory professionals when intervention might be necessary.

For example, small changes in equipment performance or testing materials could indicate an emerging problem before conventional quality-control thresholds are breached.

The researchers propose maintaining records showing which information contributed to each recommendation. Decisions involving significant operational or clinical consequences would require human approval.

Importantly, LabARIA remains a conceptual framework. The team has not demonstrated that it reduces laboratory errors in routine clinical practice.

According to Weill Cornell Medicine, the researchers plan to test the approach using historical laboratory data before considering implementation.

The proposal also acknowledges the difficulties of connecting different hospital information systems and the risk that failures emerge when AI components hand work to one another.

Compiled with AI assistance. Written and checked by a human editor before publication.