safety · 55 topics

multilingual safety signals

training capable and safe computer use agents

mme safety

bias audits detect bias but disagree on ranking

latent space refusal anchoring

mechanist ai scientific instrument mechanisms intelligence

sysadmin measuring instrumental power seeking in frontier ai

beyond a single direction

understanding annotator safety policy with interpretability

locate prevent stereotypes llm

safety for whom refusing the right subset of a topic

anthropic trained hacker opus on hackable environments

anthropic aar claude opus alignment

aletheia quest retrospective

obliterator qwen38 27b

shieldstral 1 0 3b

anthropic cuts fable 5 biology blocks by 85

mistral shieldstral 3b safety classifier

goodfires silico catches qwen3 35b endorsing drunk driving

claude opus 5 vending bench

openai sandbox escape

openai gpt red automated red teaming

global workspace llm

delta flight hit by firework at midway

2026 06 24 papers 20260622 role confusion injection

verifier tax llm agents

making claude a chemist

built to benefit everyone our plan

2026 06 06 papers 2606 04037

biodefense in the intelligence age

2026 05 29 papers 2605.27375

gpt 5 5 instant system card

refusal in language models is mediated by a single direction

deepmind ai co clinician

where the goblins came from

gpt 5 5 system card

openai safety report lead resigns

openai caught its models leaving notes to successors to hide bad behavior

openai tumbler ridge shooting lawsuits

openai agents cyberattack incident

openai disbands preparedness team

new mexico court orders meta to pay additional 567m in child safety case

openai autonomous ai safety pause

openai frontier safety blueprint

openai commitment community safety

gpt 5 5 bio bug bounty

painting the sides of railroad rails white

lib0xc

agility humanoid robot digit 5

hikers rescued gemini planning

chatgpt teen mode

openai bets on families

for robotaxis safety must be built in

chatgpt sensitive context recognition

openai trusted contact safeguard