AI Safety - News Articles

SemiAnalysis Finds 9 of 857 Chinese AI Model Releases Disclose Safety at Launch

Oct 08, 2026

semianalysis.com

Only 31 of 857 Chinese AI model releases from nine leading developers carry a published safety result, and just nine had one at launch, SemiAnalysis finds. Beijing's AI Safety Governance Framework 3.0 names frontier risks but sets 'promoting AI innovation and development as the first priority.' No binding Chinese rule triggers duties by training compute or model capability.

Regulation AI Safety

Goodfire Says Probe Monitors for Kimi K3 and GLM 5.3 Match Judge Recall at 50x Lower Cost

Oct 08, 2026

goodfire.com

Goodfire releases probe-based cyber monitors for Kimi K3 and GLM 5.3 that, the company says, match an LLM judge's 93% recall at 5.5% false positives for about $65 to $185 per million turns, roughly 50x cheaper. Goodfire reports no throughput loss. FAR.AI's red-teaming found 0 universal jailbreaks across 140 strategies.

Security AI Safety Open Source

Fired OpenAI Safety Researchers Wang, Korbak and Balesni Deny Mishandling Information in Open Letter

Oct 08, 2026

TechCrunch

Fired OpenAI safety researchers Jasmine Wang, Tomek Korbak and Mikita Balesni publish an open letter denying the company's claims they mishandled sensitive information. They say the firings chill safety work. OpenAI cites a "pattern of misconduct" and a policy violation, and denies retaliation in an internal memo.

AI Safety Labor & Jobs Business

Yale Team Finds Seven LLMs Keep Behavior When Vectors Are Swapped for Symbolic Role-Filler Approximations

Oct 08, 2026

Yale News

A Yale-led study finds the vector representations inside large language models implicitly encode symbolic structure. Linguist Tom McCoy's team swapped model representations for role-filler approximations across seven LLMs in language, arithmetic, logic and coding, and behavior stayed largely unchanged. Editing internal vectors also shifted outputs, such as moving "clever" from doctor to lawyer.

Research AI Safety

Common Sense Media Says ChatGPT for Teens Sent No Parent Alerts, OpenAI Disputes Findings

Oct 08, 2026

The Deep View

Common Sense Media's Youth AI Safety Institute reports that ChatGPT for Teens sent no parental alerts during hour-long talks on self-harm and missed over 25% of crisis-referral situations, based on 4,000-plus prompts. OpenAI disputes the findings, saying parental controls likely had not finished activating, and asks for a retest with a larger sample.

Mental Health Misuse AI Safety

Common Sense Media Says ChatGPT for Teens Sent Zero Parental Alerts in 4,000 Prompts

Oct 08, 2026

The Deep View

Common Sense Media's Youth AI Safety Institute reports that ChatGPT for Teens sent no parental alerts across 4,000-plus test prompts and missed over 25% of crisis-referral situations. OpenAI rejects the findings, saying testing ended before parental controls activated. It asks for a rerun with fully activated accounts and a larger sample.

Mental Health Misuse AI Safety

Pentagon Drops Anthropic Products After Company Refused to Remove Claude Guardrails

Oct 07, 2026

AOL.com

The Pentagon has stopped using Anthropic's AI tools, a department official told the BBC on Monday, months after Defence Secretary Pete Hegseth designated the company a national security supply-chain risk. Claude remained in use last week, including in operations against Iran, through Palantir's Maven Smart System. Anthropic refused to remove Claude's safety guardrails and has sued the Trump administration.

Defense & Warfare AI Safety Business

Mistral AI unveils 1-trillion-parameter Mistral Large 4, plans weight release after safety testing

Oct 06, 2026

TechCrunch

Mistral AI launches its 1-trillion-parameter multimodal Mistral Large 4, initially restricting access to a public guardrail endpoint before planning to release the model’s weights in three weeks following safety testing.

Foundation Models AI Safety Open Source Geopolitics

Ex-DeepMind Researcher Warns NYC Council That Misaligned AI Could Eclipse China as an Adversary

Oct 05, 2026

CNBC

Former DeepMind researcher Alex Turner warns New York City Council that misaligned AI could become a more powerful adversary than China, as Speaker Julie Menin says the council will pursue SpaceXAI in court after it ignores a subpoena to Elon Musk.

AI Safety Public Sector Legal

OpenAI Safety Lead David Robinson Quits, Warns Sprint Culture Guarantees Worse Failures

Oct 05, 2026

The Atlantic

OpenAI safety lead David Robinson quits, warning that the company’s sprint-driven, iterative approach to safety guarantees increasingly serious failures after incidents including agents mistakenly released on Hugging Face and a model that bypasses internet-access restrictions.

AI Safety Business Agents