AI research wiki
social-stream · 2026-06-09

2026-06-09-morning

Summary

The morning slot's strongest signal is Anthropic's red team publishing measured offensive-cyber numbers for its unreleased Mythos Preview model (a cluster of 2 from @logangraham): it converts public patches into working N-day exploits for a few thousand dollars and a few hours, and Anthropic is disclosing this pre-release to give defenders lead time. That lands alongside a swirl of Mythos public-release rumors (a "public version of Mythos tomorrow," a "Claude 5 fable" with stricter cyber guardrails in 48h). The other real story is the industry money narrative: OpenAI filed a confidential IPO S-1, Cursor crossed $4B annualized revenue doubling in four months, and the US government is reportedly in talks to take an equity stake in OpenAI, an idea attributed to Sam Altman. On the tooling side, Kilo Code shipped Agent Manager (parallel agents each in their own git worktree) and flagged MiniMax M3 now at half price with a 1M-token context window, while Claude's developer platform added a connector observability dashboard and an in-app submission portal. NVIDIA noted Apple is expanding Private Cloud Compute to Google Cloud on NVIDIA GPUs. The remainder is noise: an 18-tweet @brivael run of French philosophy-of-abundance and economics posts, a Tesla FSD clip, and off-topic local-news reposts.

Posts

  • Anthropic Red measures Mythos Preview's N-day exploit uplift (cluster of 2: @logangraham, @logangraham · red.anthropic.com/2026/n-days). The team found Mythos Preview is meaningfully better at developing N-days, vulnerabilities already publicly disclosed and patched on some systems, where attackers exploit the "patch gap" on machines that have not updated. Once a patch ships, "patch diffing" the pre- and post-patch binary reveals exactly what was fixed; historically that reverse-engineering was slow (WannaCry took 59 days after its patch, Citrix Bleed ~2 weeks), buying defenders time. Mythos collapses that to hours and a few thousand dollars. Anthropic publishes deliberately so the world can prepare, and @logangraham promises more defensive/blue-team work. See responsible-ai summary.
  • Mythos public-release rumors (@ns123abc). Claims a "Claude 5 fable" (a Mythos variant with stricter cyber guardrails) drops within 48 hours, echoing a separate "public version of Mythos tomorrow" rumor circulating the same morning. Unconfirmed, but consistent with the N-days disclosure timing above.
  • Kilo Code Agent Manager: worktree-isolated parallel agents + MiniMax M3 price cut (cluster of 4: @kilocode, @kilocode, @kilocode · blog). Agent Manager runs multiple coding agents at once, each in its own git worktree (separate directory, separate branch, sees only its own files) on a To Do / Doing / Done kanban board, so agents physically cannot overwrite each other's files or bleed context. The attached promo graphic shows the same prompt ("add stripe billing") running simultaneously on Claude Opus, GPT-5.5, and Claude Sonnet, each in its own worktree, so you keep whichever answer you prefer. Separately, Kilo notes MiniMax M3 now runs at half its standard input price with a 1M-token context window, already live in their model picker, part of a broad price race (Xiaomi and DeepSeek also cut). Relates to multi-agent systems.
  • OpenAI files confidential IPO S-1, and a US government equity stake is in play (cluster of 2: @ns123abc, @ns123abc · NOTUS). OpenAI announced via Rule 135 that it submitted a confidential S-1, pre-empting a leak, with timing undecided. In parallel, reporting says the Trump administration's plan to take a direct equity stake in OpenAI was news to other AI companies, and that Altman himself pitched the equity-stake idea and has lobbied for it since early 2025. Anthropic says it has not discussed any such stake.
  • Cursor crosses $4B annualized revenue (@ns123abc · Forbes). $2B in February, $3B in April, $4B last week, doubling in four months. SpaceXAI reportedly holds a $60B option to buy Cursor. A concrete data point on how fast coding-agent revenue is compounding.
  • Claude developer platform adds connector observability + in-app submission portal (cluster of 3: @ClaudeDevs, @ClaudeDevs, @ClaudeDevs · blog). Connector developers get a dashboard to track adoption (active users, tool calls, directory rank), diagnose errors and latency with per-tool breakdowns, and compare usage across Claude, Claude Code, and Cowork, plus a new in-app portal to submit MCP connectors to the vetted Connectors Directory. Platform plumbing, but signals the MCP connector ecosystem is being industrialized.
  • NVIDIA: Apple Private Cloud Compute expands to Google Cloud on NVIDIA GPUs; transaction foundation models (cluster of 3: @nvidia, @nvidia · NVIDIA blog). NVIDIA highlights Apple expanding its privacy-preserving Private Cloud Compute to Google Cloud running on NVIDIA GPUs (the backend under the new Siri / AFM 3). Separately it pushes "transaction foundation models," transformer models trained on billions of financial events for a unified view of consumer behavior, with Revolut and Mastercard cited as training on proprietary data to beat task-specific fraud/credit models.
  • Anthropic's bcherny on a year of Claude Code (@bcherny). A retrospective video with @_catwu one year after Claude Code's GA: why he uses auto mode over plan mode, how routines fix bugs before he sees them, and why he now does most coding from his phone. Color on how the team uses its own product, not a research claim.
  • Assorted chip/policy one-liners (@ns123abc and others). Jensen Huang reportedly declined Sen. Warren's invite to testify under oath about NVIDIA's China sales; Apple delayed Siri AI in the EU citing the Digital Markets Act; chatter that TSMC views Intel's EMIB packaging as a real threat. Headlines without primary sourcing in-thread.
  • @brivael French philosophy-of-abundance and economics run (18 tweets, e.g. @brivael). A long thread on space-faring civilization, Menger-vs-Marx value theory, OnlyFans power laws, an NBER "is the iPhone birth control" paper, and a SpaceXAI orbital-compute satellite repost. Off-topic for the wiki's focus; click through if interested. Skip.
  • Off-topic / non-AI (@Tesla FSD highway clip, @WHFraudTF FEMA fraud). No AI research or industry substance. Skip.