CrowdStrike's AI Kill Switch Halts Rogue Agents in Real Time

Key Takeaways

• CrowdStrike launched an AI Kill Switch in Falcon Guardian, letting administrators terminate rogue AI agents with full audit trails.
• Only 11% of organizations automatically block overstepping AI agents, while 65% had an agent-related incident in the past year.
• Utah became the first state to allow AI-only acne prescriptions through Nolla Health, with physician oversight built into the pilot.
• The OLIVIA protocol found that AI stock prediction models lose accuracy under real trading conditions and fail to survive trading costs.
• AlgoPilot improved average accuracy from 38.4% to 72.9% across 13 benchmarks by teaching language models to choose the right strategy.
• NVIDIA's Factory Operations Blueprint (FOX) uses AI agents for defect detection and runs entirely on-premises at a DGX Station.
• A former Meta employee was laid off after her performance rating dropped during medical leave, raising concerns about AI in workforce decisions.
• Nearly two-thirds of workers worry that companies use AI to select employees for layoffs, though Meta denies AI played any role.
• 76% of registered voters oppose local data center construction, creating uncertainty for AI data center investments tied to stocks like NVIDIA.
• Legal experts urge companies to verify data sources and document data acquisition practices before training AI products to avoid intellectual property issues.

CrowdStrike launches AI kill switch to stop rogue agents

CrowdStrike has launched an AI Kill Switch built into its Falcon Guardian platform. Security teams can terminate a running agent, block it from relaunching, and cut off its network access. An authorized administrator must initiate the action, and every step is recorded in an audit trail. Analysts say the tool is helpful but should not be treated as a complete fix for AI security risks.

CrowdStrike lets security teams kill rogue AI agents

CrowdStrike introduced a feature that lets administrators block an AI agent from reaching its model. The tool works through an AI gateway included with Falcon Guardian or a company's existing gateway. A survey found that only 11% of organizations automatically block an agent that oversteps, while 65% had an agent-related incident in the past year. Other companies like Microsoft and Cisco use different approaches to limit agent risk.

CrowdStrike gives security teams a kill switch for rogue AI agents

CrowdStrike added an AI Kill Switch to its Falcon Guardian platform. It lets security teams stop agents that go off-script by using the Falcon sensor on endpoints. Administrators can terminate agents at the gateway, with endpoint termination planned later. All actions need authorized approval and are logged. CEO George Kurtz said the feature helps balance AI safety with faster adoption.

AI starts writing acne prescriptions without a doctor in Utah

Utah became the first state to allow an AI program to prescribe medication without a doctor, starting with acne treatments. Nolla Health, based in New York City, began the pilot covering eight topical treatments. Patients use Nolla's app to scan their face and share medical history. A physician reviews the first 100 patients, then weekly checks, then at least 10% monthly spot-checks. The pilot costs about $5 a month for the app.

AI writes acne prescriptions without a doctor visit in Utah

Nolla Health began offering AI-generated acne prescriptions to Utah patients on Oct. 5. The AI scans skin, checks oiliness, and scores acne severity through the app. Utah set aside its usual rule requiring a licensed prescriber for this pilot. The American Medical Association warned that AI tools must meet high standards. Two other companies also received pilot approval but need more regulatory sign-off. The pilot runs for one year and may be renewed twice.

OLIVIA test shows AI stock predictions fade under real trading conditions

Researchers created the OLIVIA protocol to test if AI models can truly predict stock returns without looking ahead or memorizing data. They tested five AI models reading news about S&P 500 stocks. A fine-tuned open-weights model performed well on validation data but its signal dropped sharply on blinded test data. Commercial models only predicted the overnight price move at market open, not later moves. No model could distinguish over-reaction from under-reaction, and no strategy survived trading costs.

AlgoPilot teaches language models to pick the right problem-solving strategy

AlgoPilot is a framework that helps language models choose and structure the right approach for different types of problems like programming, reasoning, and planning. It has three parts: a SteerLM that selects the strategy, a GuideLM that creates structured guidance, and an ExecutionLM that produces solutions. Built on Qwen3-8B, it was tested across 13 benchmarks and 66 domains. Average accuracy improved from 38.4% to 72.9%, outperforming large model prompting baselines. The steering and formulation modules can transfer to other models.

Workers fear AI is being used to decide who gets laid off at companies

A former Meta employee was laid off after her performance rating dropped during medical leave, raising concerns that AI tools influenced the decision. At Meta, dashboards tracking employee performance failed to account for leave, making it impossible to detect if a ranking was compromised. A survey found nearly two-thirds of people worry about companies using AI to select employees for layoffs. Meta denies AI played any role in layoff selections. Experts say using AI for workforce decisions is still in its early stages but is becoming harder for workers to understand.

Midterm election results threaten AI data center investments

Democratic victories in midterm elections have raised concerns about the future of AI data center development. A Fox News Poll showed 76 percent of registered voters oppose local data center construction. Market strategists discussed how the election outcome could stall AI data center projects and impact major AI stocks like NVIDIA. Congressional scrutiny from Greg Casar added to the uncertainty. The discussion focused on market implications for the AI industry.

NVIDIA's Factory Operations Blueprint uses AI agents to find production defects

NVIDIA's Factory Operations Blueprint (FOX) is a reference architecture for a Factory Manager Agent that uses an LLM to investigate production issues on assembly lines. The system connects to machine databases, manufacturing execution systems, and enterprise resource planning tools through adapters. Specialist AI agents handle tasks like visual inspection using NVIDIA Cosmos and time-series anomaly detection using Kumo. The system can run fully on-premises at an NVIDIA DGX Station to keep all production data local. A reference example traces an AOI alert through inspection, rework, and root cause analysis on printed circuit board assembly lines.

Legal experts warn of AI training data risks

Technology experts advise companies to be careful when using data to train artificial intelligence products. Speed is important for business, but rushing can lead to legal problems with intellectual property laws. Companies should verify that data sources are legal and check contracts to avoid sharing restricted knowledge. It is also necessary to document how data is obtained before any issues arise. Following these steps helps ensure innovation does not violate the rights of others.

Mexican tech firms map employee AI skills

Companies in Mexico are changing how they track employee abilities by using assessment-based AI skills maps. This shift happens because the demand for AI users and developers is growing rapidly in the region. Instead of relying on self-reported profiles, businesses now use tests to identify specific skills needed for AI work. This approach helps both companies stay competitive and allows employees to develop their careers. Experts expect this method to become even more popular in the coming years.

New method speeds up discrete diffusion models

Researchers propose Bidirectional Centered Score Matching to improve how discrete diffusion models are distilled. This new technique creates an efficient objective that reduces computational complexity compared to previous methods. The approach uses a bidirectional active set to lower the number of comparisons needed during training. Experiments show the method offers strong quality and efficiency trade-offs. It achieves up to six times fewer network forward evaluations while improving performance on HumanEval and MMMU tasks.

Sources

• CrowdStrike launches AI kill switch to shut down rogue agents
• CrowdStrike Lets Security Teams Kill Rogue AI Agents
• CrowdStrike gives security teams a kill switch to stop rogue AI agents
• AI Starts Writing Acne Prescriptions Without A Doctor’s Visit In One State
• AI starts writing acne prescriptions without a doctor's visit in one state
• Thinking Fast, Trading Late: System One Decision Models vs the Market
• AlgoPilot: Cross-Paradigm Reasoning in Language Models via Strategy Selection and Guidance
• Am I going to be fired by AI? How corporate America is being seduced by AI and algorithms in the ‘forever layoffs’ era
• AI data center restrictions threaten tech investments after midterm victories
• How NVIDIA Factory Operations Blueprint Coordinates AI Agents to Investigate Production Issues
• Gavel to Gavel: AI training data
• How Companies Are Mapping AI Skills Across Tech Teams
• Bidirectional Centered Score Matching for Discrete Diffusion Distillation