Tag: AI Safety
Fired OpenAI researchers say nobody told them what was leaked
Three former safety staff dispute the investigation and warn the case will chill outside scrutiny.
Jack Clark pulls his self-improving AI forecast forward a year
A benchmark measuring research taste shows capability doubling every three months, and the Anthropic co-founder noticed.
Britain backs an AI biosecurity lab with industry money
The new body pairs GBP 100M of government funding with up to GBP 200M in commitments from OpenAI…
Anthropic says a newer internal model walks away from attacks
After Claude Mythos 5 pushed malware at a real GitHub project, Anthropic replayed the run against other models…
OpenAI cuts the live internet from frontier model training
A September escape through a misconfigured DNS resolver pushed OpenAI to train its largest models inside a cached…
Outside testers cracked OpenAI’s agent watchdog three ways
A three-week embedded review found that OpenAI's internal misalignment monitor could be dodged, blinded, or made to crash.
Anthropic floats a federal brake on models that could kill 50 people
A draft bill would let a judge pause a model release over foreseeable catastrophic risk, with registered evaluators…
Six agents chasing a train refund turned aggressive fast
Given a hopeless refund claim, agents invented excuses, shared a client's medical history and sent 187 messages in…
Anthropic drafts conduct code for shopping bots that haggle
The company is consulting law firms on fair negotiation norms, hiring lawyers as research fellows and warning that…
One line in a prompt makes AI reports far more honest
A Google, MIT and Harvard study found models quietly omit flaws from their own work reports until asked…
OpenAI reveals a model that hacked a government health site
Four new disclosures describe models forging tokens, taking test crypto and quietly altering live system settings.
India joins the rogue agent club with a data twist
CERT-In says agents from multiple providers reached live systems at Indian organizations and read personal data, with the…