Google has paused its open source bug bounty program after a flood of AI-generated submissions overwhelmed its engineers. The company stopped accepting new vulnerability reports through its Open Source Software Vulnerability Reward Program on October 1, 2026. Most of the automated submissions turned out to be invalid, clogging up both Google's review team and the open source maintainers who rely on the system. The OSS VRP, which launched in 2022 and covers projects like Go, Angular, and Protocol Buffers, will remain paused until next year, with Google planning an update in Q1 2027. Reports submitted before the pause are still valid.
In other AI infrastructure news, LinkedIn has found a way to cut GPU downtime using fault tolerance software from Clockwork.io. The social media company deployed LinkPass across its AI hardware fleet, and the tool automatically reroutes traffic around failing components. This has saved tens of thousands of GPU-hours each month from going to waste. When cables, optics, or network interface cards break, jobs keep running while repairs happen in the background.
Clockwork.io itself just raised $31 million in a funding round co-led by Seligman Ventures, Wing Ventures, and Premji Invest. The company has now pulled in $73 million total. Its software tackles a real problem in large GPU clusters: hardware failures happen daily, and recovering from snapshots can take up to 90 minutes, during which healthy GPUs sit idle repeating work that was already done. Clockwork's approach keeps those GPUs productive instead.
On the research front, two new papers address how AI systems learn and improve. One introduces Harness-Aware Distillation (HAD), a method for shrinking large language model agents into smaller ones. Instead of imitating everything the teacher model does, HAD focuses only on what the teacher adds beyond existing harness software, using action preferences and validity checks to guide learning. The result is smaller models that outperform standard distillation and get stuck in fewer unproductive loops.
A separate research effort tackles recursive self-improvement for reasoning tasks. Called task-harness co-evolution, the method generates progressively harder problems in math, coding, and science by updating both the tasks and the system that creates them. Over fourteen rounds, mean solver accuracy fell from 100% to 54.8%, showing how challenging sustained improvement really is. Still, a 27B model trained on 10,000 synthesized math examples reached 62.5% accuracy on the APEX benchmark.
Regulatory and educational developments are also shaping how AI gets used. New York's financial regulator is collaborating with California and Illinois on AI safety rules under the Responsible Artificial Intelligence Safety and Education Act, signed into law by Governor Kathy Hochul in December 2025. The law requires major AI developers to register with the state. Meanwhile, Oklahoma schools are adopting AI tools in classrooms but with guardrails. The state education department issued guidelines to balance learning benefits against concerns about cheating and excessive screen time, and some districts report mixed results so far.
Concerns about AI and safety extend beyond schools. OpenAI canceled the release of a new AI model due to safety concerns, and Microsoft co-founder Bill Gates has warned about the risks the technology poses. Experts say AI helps scammers work faster and more convincingly, and they urge people to take protective measures. The growing power of AI raises worries about catastrophic outcomes if proper safeguards are not put in place.
AI is also making its way into practical applications. The Department of Veterans Affairs is rolling out AI-powered upgrades to its pharmacy module, streamlining prescription processing based on feedback from pharmacists. Seven of twenty planned improvements are live now, with the rest coming in early 2027. VA and Oracle are developing a new module called Pharmacy Forward for that same year. Separately, CCC Intelligent Solutions is weaving AI into its automotive insurance claims platform, saying the technology enhances existing workflows and opens up new product opportunities.
On the cybersecurity side, ESET launched a new product lineup called Protect to fight AI-driven threats. The three tiers, Protect Core, Protect Premium, and Protect Ultimate, bundle in features like email security, multifactor authentication, and managed detection and response. ESET is drawing on AI security technology it has built since 1997 and pricing its products by seat.
Key Takeaways
- Google paused its open source bug bounty program on October 1, 2026, due to a surge in invalid AI-generated submissions, with no clear restart date until 2027.
- LinkedIn cut GPU downtime by deploying Clockwork.io's LinkPass fault tolerance software, saving tens of thousands of GPU-hours monthly.
- Clockwork.io raised $31 million in a new round, bringing its total funding to $73 million, to build fault tolerance tools for large GPU clusters.
- Researchers introduced Harness-Aware Distillation (HAD), a method that distills large language model agents into smaller, more efficient models by focusing on what the teacher adds beyond existing harness software.
- A new recursive harness self-improvement framework generates harder reasoning problems across math, coding, and science, with a 27B model reaching 62.5% accuracy on APEX after training on 10K synthesized examples.
- New York, California, and Illinois are collaborating on AI safety rules under the Responsible Artificial Intelligence Safety and Education Act, requiring major AI developers to register with the state.
- Oklahoma schools are adopting AI tools with new guidelines to address academic integrity and screen time concerns, reporting mixed results from early implementation.
- OpenAI canceled a new AI model release over safety concerns, and Bill Gates has warned about AI risks as scammers use the technology to work faster and more effectively.
- The Department of Veterans Affairs is adding AI tools to its pharmacy module, with seven of twenty improvements rolling out now and a new Pharmacy Forward module planned for 2027 with Oracle.
- ESET launched its Protect portfolio, including Protect Core, Protect Premium, and Protect Ultimate tiers, to combat AI-driven cyber threats using seat-based pricing.
Google pauses open source bug bounty amid surge in AI submissions
Google has paused its open source bug bounty program due to a significant rise in automated AI submissions. The company stated that most of these submissions are not valid. The pause will last until next year. Participants are encouraged to use Google's other bug bounty programs in the meantime.
AI-generated reports force Google to freeze open-source bug bounty
Google stopped accepting new vulnerability reports through its Open Source Software Vulnerability Reward Program (OSS VRP) on October 1, 2026. A flood of invalid, AI-generated submissions overwhelmed engineers and open source maintainers. The OSS VRP, launched in 2022, covers projects like Go, Angular, and Protocol Buffers. Reports submitted before October 1 are still valid. Google plans to provide an update in Q1 2027.
LinkedIn cuts GPU downtime using Clockwork.io fault-tolerance software
LinkedIn deployed LinkPass network fault tolerance from Clockwork.io across its AI infrastructure fleet. The software automatically reroutes traffic onto healthy paths, preventing tens of thousands of GPU-hours of downtime each month. When links, optics, cables, or NICs fail, jobs continue uninterrupted during repairs. This has transformed LinkedIn's operating model by reducing the impact of hardware failures on its GPU fleet.
Clockwork.io raises $31M to keep AI workloads running smoothly
Clockwork.io raised $31 million in funding co-led by Seligman Ventures, Wing Ventures, and Premji Invest. The company develops fault tolerance software for AI clusters with thousands of GPUs. Hardware failures happen daily in large clusters, and recovery from snapshots can take up to 90 minutes, leaving healthy GPUs idle. Clockwork's software keeps GPUs doing useful work instead of repeating completed tasks. The company has now raised $73 million total.
New method improves distillation for small language model agents
Researchers proposed Harness-Aware Distillation (HAD) for distilling large language model agents into smaller ones. Standard distillation imitates the teacher's full outputs, but HAD focuses only on what the teacher adds beyond the existing harness software. HAD uses an action preference and a validity check to improve learning. Across multiple benchmarks, HAD outperforms standard on-policy distillation and enters fewer unproductive loops.
New York leads AI safety efforts with California and Illinois
New York's financial regulator is working with California and Illinois on AI safety rules. Superintendent Kaitlin Asrow is enforcing the Responsible Artificial Intelligence Safety and Education Act. Governor Kathy Hochul signed the law in December 2025. The law requires major AI developers to register with the state. New York aims to lead on AI safety as federal lawmakers debate rules.
Oklahoma schools use AI with new limits for students
Oklahoma schools are adopting AI technology in classrooms but with clear limits. The state education department issued guidelines to balance AI benefits with concerns about cheating and screen time. Districts must ensure AI tools support learning and protect academic integrity. Some districts have already started using AI with mixed results. Teachers report both improved engagement and worries about student well-being.
AI helps scammers work faster, experts warn
AI has made scammers more efficient and harder to spot. OpenAI canceled the release of a new AI model over safety concerns. Tech leaders including Microsoft co-founder Bill Gates warned about AI risks. Experts urge people to take steps to better protect themselves from scams. The growing power of AI raises concerns about catastrophic outcomes if not managed carefully.
VA pharmacy adds AI tools to health records system
The Department of Veterans Affairs is adding AI-powered upgrades to its pharmacy module. The new tools streamline prescription processing and reduce steps for pharmacists. VA Deputy Secretary Paul Lawrence said the improvements were driven by pharmacist feedback. The first seven of 20 improvements are rolling out now, with more coming in early 2027. VA and Oracle are also developing a new module called Pharmacy Forward planned for 2027.
CCC Intelligent Solutions uses AI to improve insurance claims
CCC Intelligent Solutions is integrating AI into its automotive insurance claims and collision repair platform. The company says AI enhances existing workflows without disrupting customer relationships. The move creates new opportunities to sell additional products. CCC's valuation looks attractive based on EBITDA multiples. The company's AI adoption positions it well for growth in insurance technology.
ESET launches Protect portfolio to fight AI-driven cyber threats
ESET released a new cybersecurity product lineup called Protect. It includes three tiers named Protect Core, Protect Premium, and Protect Ultimate. The tiers add services like email security, multifactor authentication, and managed detection and response. ESET uses AI security technology it has developed since 1997 and charges using seat-based pricing.
New framework improves recursive self-improvement for reasoning data
Researchers introduced task-harness co-evolution, a method called recursive harness self-improvement. It generates harder reasoning problems in math, coding, and science by updating both tasks and the generation harness. Across fourteen rounds, mean solver accuracy dropped from 100.0% to 54.8%. A 27B model trained on 10K synthesized math examples reached 62.5% mean-16 accuracy on APEX.
Sources
- Google froze its open source bug bounty program due to a 'significant rise' in AI submissions
- AI slop submissions force Google to freeze its open-source bug bounty
- LinkedIn avoids tens of thousands of hours of computing downtime monthly with Clockwork.io.
- Clockwork.io bags $31M in funding to keep AI inference and training workloads running like ... clockwork
- Harness-Aware Distillation for Small Language Model Agents
- NY’s Financial Regulator Working With California, Illinois on AI
- Oklahoma schools embrace AI while setting limits for students
- AI has made scammers more efficient. How to better protect yourself against scams
- VA Pharmacy Upgrades Bring New AI Capabilities to EHRs
- CCC Intelligent Solutions leverages AI to enhance insurance claim workflows, boosting product sales potential.
- ESET launches Protect to simplify cybersecurity amid AI-driven threats
- Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis
Comments
Please log in to post a comment.