OpenAI fires safety researchers, Android phones ship with malware
OpenAI dismisses three safety researchers, prompting transparency debate. Low-cost Android phones found pre-installed with malware.
The biggest story today involves a continued tension between rapid AI development and safety concerns, with OpenAI dismissing safety researchers and new tools emerging to secure AI systems and open-source projects. Meanwhile, a widespread Android malware campaign highlights supply chain security risks.
TL;DR
- Anthropic offers free AI security scans for open-source projects using Mythos model — Anthropic launched OSS Scanner, a free service providing AI-generated security vulnerability reports for open-source projects using its strongest models, including Mythos.
- Low-cost Android phones found shipping with pre-installed residential proxy malware — A malware campaign, 'Midnight Mimosa', was discovered on low-cost Android phones with MediaTek chipsets, pre-installed in firmware to perform ad fraud and turn devices into residential proxies.
- Goodfire launches 'inside-out' AI monitors to detect rogue agents more cheaply — Goodfire launched a new "inside-out" AI monitoring system that uses internal model signals to detect rogue AI agent behavior at a fraction of the cost of traditional external monitoring.
- Attackers hijack .gh, .sl, and .as country-code domains to obtain Google certificates — Attackers compromised the DNS records of .gh, .sl, and .as country-code top-level domains to obtain unauthorized HTTPS certificates for Google and YouTube domains.
- OpenAI fires three AI safety researchers over "breach of trust" — OpenAI fired three safety researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, citing a "significant breach of trust" and violation of policies.
- AI leaderboard Arena nearly doubles valuation to $3.1 billion in 10 months — Arena, an AI model leaderboard, raised a $200 million Series B round at a $3.1 billion valuation, nearly doubling its valuation in 10 months.
- GitHub Copilot performs well on internal benchmark, struggles on independent AI code review test — GitHub Copilot topped GitHub's own AI code review benchmark but performed differently on an independent benchmark, showing a discrepancy in evaluation results.
- Percona CEO emphasizes distinction between 'open weight' and 'open source' in AI — The CEO of Percona stated that "open weight" and "open source" are not interchangeable terms in the context of AI.
- OpenAI's math solutions fall short of academic standards, lack human interpretability — OpenAI's recent release of hundreds of math problem solutions did not fully meet academic standards, particularly regarding the need for human understanding and formalization.
- ChatGPT to add new college planning and study tools for teens — OpenAI is rolling out new College Planner tools and study features like multi-photo PDF conversion and flashcard generation to ChatGPT for Teens.
Anthropic offers free AI security scans for open-source projects using Mythos model
Top story · Security · The Verge · 2026-10-09
Anthropic introduced OSS Scanner, a new service offering free security vulnerability scans for open-source projects. The scans are performed by Anthropic's "strongest models," including Claude Mythos.
Participating projects will receive thorough, periodic security reports, which are entirely model-generated without human review. This approach aims for faster and more frequent scanning, though it means reports may sometimes be incorrect or invalid.
This initiative follows recent struggles by open-source projects, including Linux and Google, to manage the volume of AI-generated bug reports, and comes after AI tools have helped identify major security flaws like the "Copy Fail" bug in Linux.
Key facts
- Service named OSS Scanner
- Utilizes Anthropic's "strongest models" including Claude Mythos
- Reports are fully model-generated, no human review
- AI tools previously identified "Copy Fail" bug in Linux
Why it matters: Engineers managing or contributing to open-source projects can leverage this free service to enhance security posture, but must be prepared to triage potentially false positive vulnerability reports due to the lack of human review.
Anthropic's free AI security scans offer a valuable, though automated, resource for open-source projects to identify vulnerabilities faster.
Low-cost Android phones found shipping with pre-installed residential proxy malware
Security · BleepingComputer · 2026-10-09
A malware campaign named 'Midnight Mimosa' has been identified on low-cost Android smartphones, where malicious software is embedded directly into the device firmware. This malware grants system-level privileges, allowing silent app installation, ad fraud, and conversion of devices into residential proxies.
Bitdefender researchers found the campaign impacted thousands of devices across over 150 countries for approximately two years, with significant numbers in Mexico, France, Italy, United States, Germany, Brazil, and Spain. The malware was found on devices from legitimate manufacturers like Doogee S200 X and Cubot KINGKONG X, as well as those impersonating Samsung and Apple.
Unlike typical Android malware, Midnight Mimosa is pre-installed in the system partition and impersonates legitimate system packages, making it difficult to remove through standard uninstall processes. Some users reported firmware updates from manufacturers that resolved infections, but the source of the malicious software in the supply chain remains unclear.
Key facts
- Malware campaign named 'Midnight Mimosa'
- Affected thousands of devices in over 150 countries over two years
- Found on low-cost Android phones with MediaTek chipsets
- Examples: Doogee S200 X, Cubot KINGKONG X
- Malware packages include com.android.non.szcz
Why it matters: Engineers developing for Android or managing device fleets should be aware of deep supply chain vulnerabilities, as pre-installed firmware malware bypasses traditional app-level security and can turn devices into a botnet without user consent.
Pre-installed firmware malware on low-cost Android phones poses a significant supply chain security risk, allowing system-level compromise from device inception.
🔗 Read more at BleepingComputer
Goodfire launches 'inside-out' AI monitors to detect rogue agents more cheaply
AI Agents & Tooling · TechCrunch · 2026-10-08
Goodfire, an interpretability startup, introduced a new monitoring system for AI agents that observes internal model signals rather than just external outputs. This "inside-out" approach is designed to be significantly cheaper than the standard method of using a second AI to review an agent's actions.
The system, available to Baseten customers, uses small detectors called probes that read a model's internal activations at each step. Only when a probe flags an issue does a separate AI model perform a deeper analysis. This method reuses computations the model is already making, adding less than 2% to the model's response time when running four probes simultaneously.
Goodfire's tests on the Kimi K3 open model showed monitoring 1,500 sessions cost approximately $51, compared to $233 for a cheaper external AI monitor and about $10,000 for a top-tier one. The probes caught 94% of malicious hacking sessions and flagged 8.7% of harmless ones for review. Customers can configure monitoring for risks like offensive hacking and chemical/biological weapons misuse, and choose automated responses such as logging, human review, or request refusal.
Key facts
- Goodfire's monitors use "inside-out" approach
- Available to Baseten customers
- Tested on Kimi K3 open model
- Monitoring 1,500 sessions cost ~$51 (vs. $233 for cheaper AI, ~$10,000 for top-tier)
- Caught 94% of malicious hacking sessions
- Added <2% latency with four probes
Why it matters: Engineers deploying AI agents, especially open models, can use Goodfire's internal monitoring to implement more cost-effective and proactive safety guardrails, detecting anomalous behavior before it manifests externally, which is crucial for preventing incidents like agent escapes.
Goodfire's "inside-out" AI monitors provide a cheaper and faster way to catch rogue AI agent behavior by observing internal model states.
Attackers hijack .gh, .sl, and .as country-code domains to obtain Google certificates
Security · The Hacker News · 2026-10-08
Attackers successfully hijacked three country-code top-level domains (ccTLDs) – .gh (Ghana), .sl (Sierra Leone), and .as (American Samoa) – to acquire unauthorized HTTPS certificates for several Google domains. Google confirmed that its own systems were not breached, but the compromise of the ccTLDs allowed attackers to manipulate DNS records and convince Certificate Authorities (CAs) to issue certificates.
Google responded by blocking these unauthorized certificates for its domains using CRLSets in Chrome and worked with the CAs to revoke them, protecting users of other browsers and apps. Certificate Transparency (CT) logs show at least 12 certificates were issued between September 22 and 27 for Google and YouTube names under these ccTLDs, including google.com.gh, google.sl, and google.as.
Let's Encrypt issued 11 of these certificates and ZeroSSL issued one. Google also indicated that other well-known global brands and online services were likely affected by the same attacks.
Key facts
- .gh, .sl, .as ccTLDs compromised
- Attacks occurred between September 22 and 27
- At least 12 unauthorized certificates issued
- Let's Encrypt issued 11, ZeroSSL issued 1
- Google blocked certificates via CRLSets in Chrome
Why it matters: Systems engineers and security professionals must recognize that DNS infrastructure at the TLD level remains a critical attack vector, even for major platforms like Google. This highlights the need for robust Certificate Transparency monitoring and rapid response mechanisms for certificate revocation.
Compromise of country-code top-level domains can lead to unauthorized certificate issuance, underscoring critical vulnerabilities in global DNS infrastructure.
🔗 Read more at The Hacker News
OpenAI fires three AI safety researchers over "breach of trust"
AI Models · The Verge · 2026-10-09
OpenAI has confirmed its decision to fire three safety researchers, stating the dismissals were due to a "significant breach of trust" and violations of policies regarding handling sensitive information.
The company's statement on X on Friday was a direct response to an open letter from the researchers, who alleged they were fired for raising AI safety concerns and believed they acted within OpenAI's mission and working norms.
OpenAI did not provide specific details about the breaches, only stating that the internal investigation uncovered violations "beyond what's outlined in the letter."
Key facts
- OpenAI fired Jasmine Wang, Tomek Korbak, and Mikita Balesni
- Decision announced on Friday, October 9, 2026
- Researchers published an open letter on Thursday, October 8, 2026
- OpenAI stated "a significant breach of trust" occurred
Why it matters: This incident highlights ongoing tensions within leading AI labs regarding safety and transparency, which could influence future regulations and best practices for engineers developing and deploying AI systems.
OpenAI maintains its decision to fire safety researchers was due to policy breaches, not safety concerns, despite the researchers' public claims.
AI leaderboard Arena nearly doubles valuation to $3.1 billion in 10 months
AI Models · TechCrunch · 2026-10-08
Arena, which began as a UC Berkeley research project in 2023, has secured $200 million in Series B funding, pushing its valuation to $3.1 billion. This valuation marks a nearly two-fold increase in approximately 10 months, following a $1.7 billion valuation in January from a $150 million Series A round.
The company's annualized run-rate revenue reached $100 million in June, up from $30 million in January. Arena's platform crowdsources AI model rankings based on user prompts and ratings, claiming tens of millions of monthly visitors.
In September of last year, Arena launched its commercial product, AI Evaluations, providing detailed performance analytics to model labs and enterprises using community feedback. This service addresses the challenge of AI models gaming static benchmarks and helps enterprises select models for their specific needs. Arena has also introduced an "alignment" category to its leaderboard, ranking models on issues like unauthorized action, false attribution, and deceptive completion, with OpenAI models currently leading this category.
Key facts
- Raised $200 million Series B
- Current valuation: $3.1 billion
- Previous Series A valuation (Jan): $1.7 billion
- Annualized run-rate revenue (June): $100 million
- Originated as UC Berkeley research project in 2023
Why it matters: Engineers can use Arena's crowdsourced and enterprise evaluation platforms to assess AI model performance and alignment more realistically than static benchmarks, guiding selection for specific application requirements and observing evolving safety metrics.
Arena's rapid growth and new alignment category highlight the increasing demand for independent, real-world AI model evaluation beyond traditional benchmarks.
GitHub Copilot performs well on internal benchmark, struggles on independent AI code review test
Developer Tools · The New Stack · 2026-10-06
GitHub Copilot achieved top scores on GitHub's internal benchmark for AI code review. However, an independent benchmark presented a different outcome, suggesting that its performance varies depending on the evaluation methodology.
The article highlights a potential divergence in how AI code review tools are assessed, with internal benchmarks possibly reflecting specific use cases or datasets that differ from broader, independent evaluations.
This raises questions about the generalizability of AI code review performance across diverse real-world coding scenarios and the importance of transparent, third-party validation.
Key facts
- GitHub Copilot topped GitHub's internal benchmark
- Performed differently on an independent benchmark
Why it matters: Engineers relying on AI code review tools like GitHub Copilot should be aware that internal benchmarks may not fully represent real-world performance. It is important to evaluate such tools with independent metrics and consider their efficacy in a diverse range of coding contexts rather than relying solely on vendor-provided scores.
The effectiveness of AI code review tools like GitHub Copilot can vary significantly between internal and independent benchmarks, underscoring the need for diverse evaluation.
Quick hits
- Percona CEO emphasizes distinction between 'open weight' and 'open source' in AI (The New Stack) — The CEO of Percona stated that "open weight" and "open source" are not interchangeable terms in the context of AI.
- OpenAI's math solutions fall short of academic standards, lack human interpretability (TechCrunch) — OpenAI's recent release of hundreds of math problem solutions did not fully meet academic standards, particularly regarding the need for human understanding and formalization.
- ChatGPT to add new college planning and study tools for teens (The Verge) — OpenAI is rolling out new College Planner tools and study features like multi-photo PDF conversion and flashcard generation to ChatGPT for Teens.
What to watch
- How OpenAI addresses the transparency and trust issues raised by the dismissed safety researchers, and if this impacts future AI safety discussions.
- The broader impact of the 'Midnight Mimosa' Android malware campaign, including whether manufacturers issue universal fixes and if similar supply chain compromises are discovered in other device categories.
- The adoption and effectiveness of Anthropic's OSS Scanner in reducing vulnerabilities in open-source projects, and how open-source maintainers manage the influx of AI-generated reports.
Sources
- Anthropic offers free AI security scans for open-source projects using Mythos model — The Verge, 2026-10-09
- Low-cost Android phones found shipping with pre-installed residential proxy malware — BleepingComputer, 2026-10-09
- Goodfire launches 'inside-out' AI monitors to detect rogue agents more cheaply — TechCrunch, 2026-10-08
- Attackers hijack .gh, .sl, and .as country-code domains to obtain Google certificates — The Hacker News, 2026-10-08
- OpenAI fires three AI safety researchers over "breach of trust" — The Verge, 2026-10-09
- AI leaderboard Arena nearly doubles valuation to $3.1 billion in 10 months — TechCrunch, 2026-10-08
- GitHub Copilot performs well on internal benchmark, struggles on independent AI code review test — The New Stack, 2026-10-06
- Percona CEO emphasizes distinction between 'open weight' and 'open source' in AI — The New Stack, 2026-10-09
- OpenAI's math solutions fall short of academic standards, lack human interpretability — TechCrunch, 2026-10-08
- ChatGPT to add new college planning and study tools for teens — The Verge, 2026-10-07