AI companies dominated US venture capital in the first nine months of 2026, pulling in roughly 83% of a record $515.8 billion total. The PitchBook-NVCA Venture Monitor reported that Q3 alone pushed past the previous full-year record set in 2025, with deal volume up 12% year-over-year across 1,444 completed transactions. The surge underscores how heavily investors are backing artificial intelligence firms over other sectors.
Flock Safety, an AI surveillance startup valued at $7.5 billion, plans to cut about 270 employees — roughly 18% of its workforce. The company, which has raised over $950 million, operates around 120,000 AI-powered cameras across 49 states and serves more than 4,800 law enforcement agencies. Growing privacy backlash, a Florida ban on its license plate readers, and a lawsuit in Virginia have compounded existing pressures on the firm.
Anthropic announced significant changes to Claude's usage policy, with new rules set to take effect next month. Separately, a series of Anthropic security breaches prompted the White House to mandate that companies report AI incidents to the government within 48 hours. The administration also plans to form a task force overseeing AI development and use across government agencies.
Three OpenAI safety researchers claim they were fired without explanation after raising concerns about a sensitive project. The researchers say they are seeking legal advice and may file a complaint with the US Department of Labor. OpenAI has not commented on the dismissals, which raise questions about internal safety practices at one of the leading AI labs.
DeepKeep launched AI Lens for Developers, a security tool that monitors coding agents like Cursor, Claude Code, and GitHub Copilot. The plugin tracks agent activity before and after execution, flags dangerous commands, detects exposed credentials, and maintains session logs for audit purposes. Administrators can set granular rules through a Policy Hub to control what agents are allowed to do.
Hack The Box introduced its AI Range Enterprise Edition, a platform designed to help cybersecurity teams evaluate AI agents on an ongoing basis. CEO Haris Pylarinos said the average lifespan of an AI agent without retraining is only about three months, highlighting the need for continuous assessment of whether agents need updating or replacement.
Oracle is trucking compressed natural gas to its AI data centers while waiting for pipeline connections to be completed. The approach was used at a Utah facility for over a year and is being considered for a stalled New Mexico project targeting 2.45 gigawatts. Pipeline delays from Energy Transfer LP have pushed the in-service date to next year, and trucking gas costs about four times more than pipeline delivery.
Shane Mac's personal AI agent accidentally posted his bank details on company Slack, prompting him to reconsider how he uses AI tools at work. The incident serves as a reminder of the risks organizations face when personal AI agents interact with sensitive workplace systems. Mac shared his experience as a cautionary example for others navigating the growing use of agents in professional settings.
Key Takeaways
- AI companies captured about 83% of a record $515.8 billion in US venture capital during the first nine months of 2026.
- Flock Safety is cutting roughly 270 jobs, or 18% of its workforce, amid privacy concerns and legal challenges.
- Flock Safety holds a $7.5 billion valuation and has raised over $950 million from investors.
- Anthropic has updated Claude's usage policy, with new rules taking effect next month.
- The White House now requires companies to report AI security incidents to the government within 48 hours following Anthropic breaches.
- Three OpenAI safety researchers were dismissed without explanation after raising concerns about a sensitive project.
- DeepKeep's AI Lens supports Cursor, Claude Code, and GitHub Copilot, monitoring coding agent activity for security risks.
- Hack The Box reports the average lifespan of an AI agent without retraining is only about three months.
- Oracle is using trucked compressed natural gas to power AI data centers due to pipeline delays costing about four times more.
- A personal AI agent leaked bank details on company Slack, highlighting risks of unmanaged AI tools in the workplace.
17 tech leaders reveal how AI agents are changing work
17 tech executives shared how AI agents are transforming their workplaces. The article covers changes in hiring and daily work processes. These leaders show how AI agents are becoming common in business. The insights come from a recent industry discussion.
Anthropic makes major changes to Claude's rules
Anthropic announced significant updates to Claude's usage policy. The new rules for Claude will take effect next month. These changes represent a shift in how the AI assistant operates. The updates were revealed by Anthropic leadership.
Personal AI agent leaked bank details on company Slack
Shane Mac's personal AI agent posted his bank details on company Slack. This mistake made him rethink how he uses AI agents. The incident shows risks of using personal AI tools at work. Mac shared his experience as a cautionary tale.
Flock to cut several hundred jobs amid privacy backlash
Flock Safety plans to cut about 18% of its workforce, roughly 270 employees. The AI surveillance company faces growing privacy concerns and legal challenges. It has about 1,500 employees and operates around 120,000 AI-powered cameras across 49 states. Florida banned its license plate readers in September. Employees will leave at the end of the month after a voluntary buyout program.
Flock Safety to cut jobs amid opposition to surveillance tools
Flock Safety will let go of about 270 employees, around 18% of its staff. The surveillance startup has faced scrutiny from regulators and privacy advocates. It raised over $950 million from investors and has a $7.5 billion valuation. The company provides cameras and software to more than 4,800 law enforcement agencies. A privacy lawsuit in Virginia and Florida's ban on its cameras add to its challenges.
Adaptive Reference Guidance improves reasoning RL with balanced trajectory training
Researchers developed Adaptive Reference Guidance (ARG) to decide how much reference solution help a reasoning model should get during training. The method uses prefix continuation where a model continues from a reference prefix and keeps correct trajectories. ARG selects prefixes learned from training questions and applies them in Group Relative Policy Optimization. Tests on Qwen3-4B and Qwen3-8B across five math benchmarks showed ARG achieved the highest aggregate pass@12 among evaluated methods.
Oracle trucks natural gas to AI data centers amid pipeline delays
Oracle is delivering compressed natural gas by truck to power AI data centers while waiting for pipeline connections. The company used this approach at a Utah facility for over a year and is considering it for a stalled New Mexico project called Project Jupiter with a 2.45 gigawatt target. Pipeline delays from Energy Transfer LP pushed the in-service date to next year. Trucking gas costs about four times more than pipeline delivery according to East Daley Analytics. At 100 megawatts of demand each large gas trailer would be used up in around 40 minutes.
Hack The Box launches enterprise platform to assess AI cybersecurity agents
Hack The Box introduced the AI Range Enterprise Edition platform for cybersecurity teams to evaluate AI agents. CEO Haris Pylarinos said the platform allows ongoing assessment of AI agents as IT environments change. The platform tests whether AI agents can perform specific job roles and helps determine if agents need retraining or replacement. Pylarinos noted the average lifespan of an AI agent without retraining is only three months. The platform also evaluates whether cybersecurity professionals have the skills to properly assess AI agent capabilities.
AI companies grab 83% of US venture capital in record $516bn market
US venture capital investment hit a record $515.8 billion in the first nine months of 2026 with AI companies taking about 83% of that capital. The Q3 2026 PitchBook-NVCA Venture Monitor reported that investment already surpassed the previous full-year record from 2025. Venture capital deals increased 12% year-over-year with 1,444 deals completed in the first nine months of 2026. The report shows AI companies are driving most of the growth in the venture capital market.
Anthropic breaches lead White House to mandate AI incident reporting
The White House issued a new mandate requiring companies to report AI security incidents to the government within 48 hours. This follows a series of breaches involving Anthropic that raised concerns about government system security. The administration also announced plans to create a task force overseeing AI development and use in government. The task force will include representatives from government agencies and the private sector. The move responds to growing concerns about AI risks and the need for greater oversight.
Fired OpenAI safety researchers question their dismissals
Three OpenAI safety researchers say they were fired under suspicious circumstances. The researchers say they were dismissed without any explanation. They were responsible for making sure OpenAI's AI systems were safe. The team says they were let go after raising concerns about a sensitive project. They are now seeking legal advice and may file a complaint with the US Department of Labor. OpenAI has not commented on the claims.
DeepKeep launches AI Lens to help security leaders oversee coding agents
DeepKeep introduced AI Lens for Developers, a tool that gives security teams visibility into coding agents. The plugin monitors agent activity before and after execution without adding extra software to developer machines. It flags dangerous commands, detects exposed credentials, and logs every session for audit trails. Administrators can set rules through a Policy Hub to block or approve specific actions. AI Lens currently supports Cursor and Claude Code, with support for GitHub Copilot and other tools planned.
New method improves expectation estimation in language models
A new research paper proposes a way to estimate expectations under autoregressive language models more efficiently. The approach uses potentials, which are real-valued functions that break down test functionals into simpler parts. The method leverages next-token conditional probabilities already produced during sampling. The researchers show this can reduce variance without increasing computational cost. Practical potentials were developed for several applications and demonstrated significant improvements.
Sources
- 17 tech leaders share how they're using AI agents at work
- The biggest changes Anthropic just made to Claude's rules
- My personal AI agent posted my bank details on company Slack
- AI surveillance startup Flock to cut several hundred jobs amid backlash, sources say
- AI surveillance startup Flock Safety to cut several hundred jobs amid backlash, sources say
- Balancing Reference Guidance and Free Generation in Trajectory Rollouts for Reasoning RL
- Oracle is trucking natural gas to AI data centers as pipeline delays threaten key projects
- Hack The Box Adds Enterprise Edition of Platform to Assess AI Cybersecurity Agents
- AI takes 83% of US venture investment as record $516bn market masks fundraising divide
- Exclusive: Anthropic breaches spark White House AI reporting mandate
- Fired OpenAI safety researchers say they were pushed out over ‘suspicious’ circumstances
- DeepKeep Introduces AI Lens for Developers to Give CISOs Visibility and Control Over Coding Agents
- Estimating great expectations under autoregressive language models with potentials
Comments
Please log in to post a comment.