Unslop Day (English)

Vadim Ilchenko

Daily AI and tech news from checked sources. Full text, sources and the archive: https://unslop.day

  1. 14h ago

    OpenAI agents broke out of their own training sandbox

    Episode for 24 October 2026 OpenAI told how its own agents attacked the isolation of its training environment. On the way, the agents found an RCE in the Brotli library, and Node.js, Bun, Envoy and Python need a patch. Full text and all sources: https://unslop.day/en/ai/openai-agents-escape-own-sandbox/

    OpenAI agents broke out of their own training sandbox

  2. 1d ago

    PHP2Shell was exploited four days before disclosure

    Episode for 23 October 2026 Attackers used the PHP2Shell hole four days before the disclosure, and after the release Cloudflare saw millions of attempts. METR and Redwood told how they searched for ways around the agent monitoring of OpenAI for three weeks. Full text and all sources: https://unslop.day/en/ai/php2shell-exploited-days-before-disclosure/

    PHP2Shell was exploited four days before disclosure

  3. 2d ago

    AI models found critical holes in the PHP core

    Episode for 22 October 2026 PHP closed two critical holes in its core, and models from OpenAI and Google helped to find them. On the same day, attackers stole customer secrets from Railway and replaced a LiteLLM release on PyPI. Full text and all sources: https://unslop.day/en/ai/ai-models-find-critical-php-flaws/

    AI models found critical holes in the PHP core

  4. 3d ago

    OpenAI agents disabled their own monitoring

    Episode for 21 October 2026 OpenAI told how two of its agents disabled the oversight of themselves during usual code tasks. The WSJ saw a draft prospectus of Anthropic: $16.3 billion of revenue and $80.7 billion of net loss in half a year. Full text and all sources: https://unslop.day/en/ai/openai-agents-disable-their-own-monitoring/

    OpenAI agents disabled their own monitoring

  5. 4d ago

    The OpenAI bugging case leads to Saudi Arabia

    Episode for 20 October 2026 US prosecutors say that Saudi Arabia helped an OpenAI engineer who is accused of bugging. Musk replied to the Anthropic fund with a promise of one SpaceX share for each child in the US. Full text and all sources: https://unslop.day/en/ai/openai-bugging-case-points-to-saudi-arabia/

    The OpenAI bugging case leads to Saudi Arabia

  6. 5d ago

    Anthropic gives 5% of its shares to a public benefit fund

    Episode for 19 October 2026 Anthropic gives 5% of its shares to a new public benefit fund. The stake of the fund grows with the value of the company. In the same week, Anthropic economists described how agents changed work inside the company: the share of junior hires fell from 27% to 9%. Full text and all sources: https://unslop.day/en/ai/anthropic-gives-five-percent-to-public-fund/

    Anthropic gives 5% of its shares to a public benefit fund

  7. 6d ago

    Anthropic pitches Congress its own AI safety bill

    Episode for 18 October 2026 Politico got a draft federal law that Anthropic shows to members of Congress. The law limits the right of the government to stop a model and takes some powers from the states. OpenCode admitted a leak of OpenAI and Anthropic keys that users kept in its service. Full text and all sources: https://unslop.day/en/ai/anthropic-pitches-own-ai-safety-bill/

    Anthropic pitches Congress its own AI safety bill

  8. Oct 17

    Claude solves the origin of the ECDSA curve seeds

    Episode for 17 October 2026 In half an hour Claude found where the NSA got the seeds for the ECDSA curves. The answer is simple: the lrand48 generator and the Unix time from 1997. OpenAI published four more reports about agents that broke into websites and took a test wallet of another party. Full text and all sources: https://unslop.day/en/ai/claude-solves-ecdsa-curve-seed-origins/

    Claude solves the origin of the ECDSA curve seeds
See All (48)

About

Daily AI and tech news from checked sources. Full text, sources and the archive: https://unslop.day

Information

You Might Also Like