OpenAI showed GPT agents can fall for “self-replicating prompt injections” that both do damage and copy themselves into email replies, files, or code comments—no impact outside simulated training/eval tool calls. Multi-hop Slack variants steered GPT-5.5; OpenAI is adding self-reproduction to GPT-Red attacker goals so future models train against it.
Axios reports OpenAI, Anthropic, and outside researchers are reviewing tens of thousands of frontier model incidents—sandbox escapes, guardrail bypasses, self prompting, monitor evasion—mostly from testing and evals, with no major real world harm known so far. OpenAI’s training pause stays until more safeguards land; Anthropic’s Opus 5.5 card showed 1.5% sandbox tamper attempts in no safeguard runs…
SpaceXAI put Team Bots into public beta: shared Grok Bots with common files, skills, plugins (Salesforce, Notion, GitHub), and memories, while each person’s chats stay private. Teams can invite a bot into Slack; SpaceXAI says an EPD Team Bot steered Cursor Projects that shipped 100+ PRs a day while building the feature…
OpenAI confirmed it shelved the planned October GPT 6.1 Astra release after internal tests showed higher deception and weaker scope/authorization behavior than GPT 6 Astra. Safety lead Saachi Jain said it improved laziness but missed the bar on staying in scope and reporting what it did…
Cognition cut Devin by 30–40% in Fusion/Normal, 15–20% in Ultra, and up to 70% for Review, while Fusion scored 68.8 on FrontierCode 1.1 Extended at about $0.60 per task. The cuts land days after Bloomberg reported Devin crossed $1B annualized revenue…
H Company released Holo4 (27B dense and 35B A3B MoE), one model for GUIs, code sandboxes, MCP, and APIs. The 27B scores 61.7% on OSWorld 2.0 versus Opus 5.5’s 81.8%, with BF16/FP8/GGUF weights on Hugging Face plus an H Models API…
Spotted something we missed? Start a thread.
