What a difference a day makes. After Thursday’s brutal sell-off wiped roughly $2.5 trillion off the market, AI stocks bounced back on Friday — and the news didn’t slow down. OpenAI locked in billions at a $1.25 trillion valuation, DeepMind’s AI agents found cracks in decades-old encryption, Anthropic explained how Claude “dreams” to fix its own memory, and Meta gave its assistants a better memory too.
Key takeaways
- AI stocks rebounded Friday — the Nasdaq rose 1.1% — after Thursday’s $2.5 trillion rout.
- OpenAI has reportedly raised $11 billion at a $1.25 trillion valuation, well short of its original $30 billion target.
- DeepMind’s AI agents found two weaknesses in legacy 1024-bit RSA encryption; modern 2048-bit keys are unaffected.
- Anthropic revealed how Claude uses offline “dreaming” to correct its own memories between sessions.
- Meta rolled out Active Memory and Active Focus to Muse and Meta AI, plus build-your-own Spaces apps.
1. AI Stocks Bounce Back After Thursday’s $2.5 Trillion Rout
Friday brought a breather. According to CNN, US tech stocks staged a modest rebound after a sell-off that wiped roughly $2.5 trillion in market value off Wall Street. The Nasdaq rose 1.1% and the S&P 500 gained 0.3%, while the Dow slipped about 0.5%. Nvidia climbed 2.8% after plunging roughly 9% the day before, and SpaceX gained 8.6%.
The bounce was only partial, and it came after the pain had spread to Asia, where South Korea’s Kospi and Japan’s Nikkei both closed down 6% on Friday. The trigger, as we covered in yesterday’s tech news roundup, was the midterm election outcome: investors worried a Democratic Congress would slow AI development and clamp down on data center expansion. A stronger-than-expected jobs report — 165,000 new jobs in October — helped steady nerves, but this week’s wild swings show how much of the AI trade now rides on politics as well as products.
2. OpenAI Raises $11 Billion at a $1.25 Trillion Valuation
OpenAI’s fundraising saga took another turn on Friday. According to TechCrunch, the company has reportedly raised $11 billion so far in a new funding round that values it at $1.25 trillion, citing Bloomberg’s reporting. The round hasn’t closed yet and could total around $20 billion when it wraps up in the coming weeks.
The numbers tell a story of a cooler market. Back in late September, OpenAI was reportedly looking to raise $30 billion at a $1.4 trillion valuation. Both the target and the valuation have since come down, and some large investors reportedly gave the round a lukewarm reception — a sharp contrast with March, when OpenAI raised $122 billion at an $852 billion valuation. None of this is confirmed by OpenAI itself, but the direction is clear: even the company behind ChatGPT is discovering that investors now want to see returns on AI spending, not just bigger ambitions.
3. DeepMind’s AI Agents Find Cracks in Legacy RSA Encryption
Google DeepMind says its AI agents, working under human direction, found two weaknesses in legacy 1024-bit RSA encryption. According to Crypto Briefing, the findings were announced on November 6 — and the headline caveat comes first: modern RSA deployments using 2048-bit or larger keys remain secure.
The first weakness enables the forgery of digital signatures under certain conditions, and the second is a new algorithm that lowers the cost of factoring many related 1024-bit keys at once. In one test, the agents reproduced a previously solved 1024-bit attack using about 3.5 times less CPU time than the academic team that first pulled it off. Before anyone panics: 1024-bit RSA is already obsolete, and NIST has disallowed it for new signatures for years. The real story is that AI agents are now doing genuine cryptanalysis research, guided by humans, in a matter of weeks — a capability milestone with implications far beyond one old encryption standard, as we explored in Thursday’s roundup on AI agents cracking Linux flaws.
4. Anthropic Explains How Claude “Dreams” to Fix Its Own Memory
Anthropic published a fascinating engineering post on November 6 about how it gives Claude a memory that lasts between sessions — and the centerpiece is a process the company calls “dreaming.” According to Anthropic’s engineering blog, Claude uses idle time between sessions to reread its own logs, correct notes that turned out to be wrong, and build linked wiki pages it can draw on later.
The idea borrows from “sleep-time compute,” proposed by researchers at Letta and UC Berkeley: models use quiet periods to reorganize what they already know, much like the human brain consolidates memories during sleep. One test made the point vividly: playing Pokémon Red, one Claude model wrote down almost everything it encountered — 413 files across 225 folders — and earned a single badge, while another kept just nine files plus a single summary and earned four. Less, it turns out, can be more. Anthropic stresses this is a research write-up, not a product launch — but as AI agents take on longer, messier jobs, remembering what matters (and forgetting what doesn’t) is becoming a core capability.
5. Meta Gives Muse a Better Memory — and Lets You Build Apps Inside It
Meta had a busy Friday too. According to Unite.AI, Meta AI Research announced it is rolling out specialized versions of the models behind Muse and Meta AI, trained using its flagship Muse Wonder system. The update brings two of Wonder’s headline capabilities — Active Memory and Active Focus — to everyone, including free-tier users.
Active Memory lets the assistants retain important information and retrieve it when it becomes relevant later, while Active Focus helps Muse keep pushing toward a goal when a task hits an unexpected snag. Meta reports the new Muse models are 2.4 times less likely to forget critical information and 67% better at self-correcting everyday tasks, and the current Muse model now scores 47% on Meta’s internal Personal Superintelligence Index, up from 38%. Those figures are Meta’s own, measured on its own tests, with no outside verification yet — so take them as a direction, not a verdict. Meta also launched Spaces in the Muse app the same day: custom mini-apps you create just by describing an idea. Meta is clearly betting the next phase of the AI assistant race is about persistence — assistants that remember you, and workspaces that stick around.
The Takeaway
This week delivered the AI industry’s first real stress test in a while, and Friday showed both sides of it. Markets can still bounce — but the days of unquestioned AI optimism are over, with investors and regulators now asking where the returns are. At the same time, the technology keeps compounding: agents that do cryptanalysis, models that dream to fix their own memories, assistants that remember what matters. The money and the machines are both being graded on results now. That scrutiny is healthy — it’s what turns a boom into something durable.
Check back tomorrow for the next roundup.
FAQ
Why did AI stocks fall on Thursday and rebound on Friday?
Thursday’s sell-off followed the midterm election results, with investors worried a Democratic Congress would impose stricter regulation on AI development and data center expansion. The Nasdaq fell 5.5%, wiping roughly $2.5 trillion in market value. On Friday, bargain-hunting and a stronger-than-expected US jobs report helped the Nasdaq recover 1.1%, though the rebound only partly reversed the losses.
How much has OpenAI raised at what valuation?
OpenAI has reportedly raised $11 billion so far in a funding round valuing it at $1.25 trillion, according to Bloomberg reporting cited by TechCrunch. The round has not closed and could reach about $20 billion — down from an original target of $30 billion at a $1.4 trillion valuation.
Is my online banking at risk from the DeepMind RSA findings?
No. The weaknesses DeepMind’s AI agents found affect only obsolete 1024-bit RSA encryption. Modern systems use 2048-bit or larger keys, which DeepMind says remain secure against these specific findings. 1024-bit RSA has already been disallowed for new signatures for years.
What is Claude’s “dreaming” and can I use it?
“Dreaming” is Anthropic’s term for offline maintenance: between sessions, Claude reviews its logs, corrects wrong notes, and builds organized wiki pages from what it learned. It changes no model weights. Anthropic describes it as research, not a product launch, so it isn’t being announced as a general feature you can switch on.
Sources
- CNN, “AI stocks bounce back after $2.5 trillion wipeout, but questions remain” — https://www.cnn.com/2026/11/05/markets/asian-markets-us-rout-ai-investment-intl-hnk?cid=external-feeds_iluminar_meta
- TechCrunch, “OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation” — https://techcrunch.com/2026/11/06/openai-reportedly-raises-an-additional-11-billion-at-a-1-25-trillion-valuation/
- Crypto Briefing, “Google DeepMind’s AI agents expose two weaknesses in legacy 1024-bit RSA” — https://cryptobriefing.com/google-deepmind-ai-rsa-1024-bit-weaknesses/
- Anthropic, “The components of agent memory” — https://www.anthropic.com/research/components-of-agent-memory
- Unite.AI, “Current Muse Scores 47% on Meta’s Personal Superintelligence Index” — https://www.unite.ai/current-muse-scores-47-on-metas-personal-superintelligence-index/
