Synthesis
AI and security: the hosts' threat model
How the hosts expect AI models to break software and institutional security, whether labs become defence companies, and who builds the response.
Across about 18 episodes the hosts return to one worry: models that find and exploit vulnerabilities make existing security fragile. Sam Lessin predicts regional banks will be hacked and says you do not need superintelligence for chaos; Dave Morin relays that models can reverse-engineer binaries, so closed source may be ending (secondhand in two episodes). Sam’s longer thesis, which Dave endorses, is that Anthropic and OpenAI become valuable mainly as defence companies, holding unreleased ‘zero-day’ models in reserve and being paid billions to fix a subverted bank; he separately calls the stealth-bomber version a thought experiment. Dave expects a new security industry, as the virus era produced antivirus; an unnamed host doubts a cyber startup is venture-scale. Incident stories get cynical readings (OpenAI’s sandbox escape, the Mythos leak) that Jess and Dave partly resist. Reported news on chips, export controls and Pentagon disputes sits alongside. Dave co-founded a home-security company, Sauron, and Sam’s own Instagram account was hijacked; both are flagged stakes. Most claims are predictions, not evidence.
Key points
- Models make existing security fragile: regional banks hacked, binaries reverse-engineered so no closed source, document security ‘does not exist’ — 8 episodes. (Sam · The Real Reason Elon Wants Sp (Oct 2026) ▶ 03:12)
- Reported chips, export-control and national-security stories: H200 sales to China, smuggled Blackwell chips, foreign-national access, Pentagon and Anthropic terms — 6 episodes. (Jess · Founders Beware: Illiquid Valu (Jan 2026) ▶ 06:42) (Amir Efrati · Amir Efrati (May 2026) ▶ 12:26)
- A new AI-security industry and practices are expected (sandboxes, harnesses, labs red-teaming each other), though one host doubts a security startup can be venture-scale — 4 episodes. (Dave · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 20:54)
- Incident stories invite cynical reads (marketing, stock support) that others contest — 3 episodes. (Brit · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 09:35)
- Agents leak: Meta AI support was exploited for four months, offboarding now means many connectors, and a rogue-agent story was anthropomorphised — 3 episodes. (Sam · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 25:07)
- Restrictions on one service are futile while another will do it; the fast crowd wins and a slower secure core commands a premium — 2 episodes. (Sam · OpenAI’s $852 Billion Valuatio (Apr 2026) ▶ 12:31)
- Sam: labs become valuable as defence companies holding zero-day models in reserve; Dave agrees — 2 episodes. (Sam · Anthropic Wants to Slow Down. (Sep 2026) ▶ 18:52)
Where guests disagree
Was OpenAI’s reported sandbox escape real or a stunt? ⚠️
- Cynical: Brit says OpenAI was jealous of Mythos’s scary moment; an unidentified host says it supports the stock (Brit · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 09:35)
- Plausible: Jess notes Hugging Face disclosed it itself; Dave says it is totally plausible (Jess · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 10:19)
Is a cybersecurity company buildable in 2026? ⚠️
- Yes: Dave expects a new set of security companies (Dave · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 20:54)
- Doubtful: an unnamed host says the next model secures things and a security firm is mostly professional services (one of the hosts · The Real Reason Elon Wants Sp (Oct 2026) ▶ 08:58)
Frameworks
- Labs as stealth bombers: zero-day models held in reserve — Sam’s thesis for how the model labs become valuable over 5 to 10 years. (Sam · Anthropic Wants to Slow Down. (Sep 2026) ▶ 18:52)
- They cannot make money competing in the token marketplace
- With about 10,000 PhDs they develop zero-day models against other LLMs and never release them
- Like nuclear weapons: when a country is subverted, only one or two parties can fix it, and they can charge huge sums
- Zero-day only works while the model is unexposed, as exposed models get built on
- Labs as stealth bombers: zero-day models held in reserve — Sam’s thought experiment, not a pitched business model: open-source models serve almost all everyday agent work cheaply, while a few labs keep unseen, very expensive models and capacity that governments and companies call in when a bot rebellion breaks out. (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 32:22)
- Everyone uses cheap open-source models for nearly all work
- Frontier labs lose that commodity work on cost
- Labs keep secret, extremely powerful models and on-demand compute
- When a rebellion or incident hits, the labs are the only ones to call and can charge enormous sums (he floats a hundred billion dollars)
- The two internet disasters of AI — Sam’s view of how AI breaks the old internet in two stages. (Sam · OpenAI’s $852 Billion Valuatio (Apr 2026) ▶ 13:31)
- First: bye-bye web, ad models and the old business model
- Second: agents running around including things they don’t understand and checking APIs; hidden problems; no one trusts anything and the whole thing collapses
- Fast crowd versus secure core — Sam’s model of the coming split in software trust. (Sam · OpenAI’s $852 Billion Valuatio (Apr 2026) ▶ 12:42)
- The fast crowd skips checks and wins on speed
- A secure core where someone really checked all the packages charges a premium
- It will be slower and about 10 times more expensive and most people won’t care
Tactics
- Contain a model inside a sandbox or harness so it does the job without escaping. — Dave says much of harness engineering right now is simply keeping models inside a defined sandbox; he adds that his OpenClaw community is working on it. (Dave · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 21:03)
- Cut off a runaway model’s compute as a circuit breaker. — Jess, self-described as non-technical, notes the OpenAI post called the escape very compute-intensive and suggests cutting the compute; another host quips about a circuit breaker and she concedes it is probably not that simple. (Jess · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 24:46)
- Have big labs red-team each other’s models before release — As each new model is about to ship, other labs, in the US or China, test it, as security organisations already do with shared bug lists. (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 16:22)
Stories & examples
- Slow’s Instagram account is hijacked via Meta AI support: Hackers had been getting Meta AI’s customer support chatbot to change the email on big Instagram handles for four months until it was noticed that day; Slow’s account was taken and made private by a stranger. Lesson: Even huge companies let AI roam where it shouldn’t; Sam finds it funny and expects to recover the account. (Sam · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 25:07)
- Sublime Security and collaboration in security: Sam says the top ten tech companies’ security chiefs have unofficial weekly calls, and Sublime’s pitch was a collaborative security platform for everyone else. Lesson: Collaboration platforms in security could be a model for AI safety; Sam adds it will not be humans because things move too fast. (Sam · Anthropic Wants to Slow Down. (Sep 2026) ▶ 17:48)
- Unrestricted image generation on a rented GPU: Sam hit content moderation on an image API, asked Claude, which suggested booting an H100 for about two dollars an hour and downloading the software; it worked. Lesson: No single service can enforce rules when anyone can rent raw compute. (Sam · OpenAI’s $852 Billion Valuatio (Apr 2026) ▶ 26:26)
Numbers
- Hypothetical fee for a lab to put down a bot rebellion in Sam’s thought experiment: 100 billion dollars (host’s framing, approx.) (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 33:32)
- PhDs available to the labs to build zero-day models: 10000 PhDs (host’s framing, approx.) (Sam · Anthropic Wants to Slow Down. (Sep 2026) ▶ 19:05)
- Hypothetical fee to fix a subverted bank: 10 billion dollars (host’s framing, approx.) (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 20:38)
- Relayed annual Chinese social engineering budget against the US: 80000000000 dollars/year / year (host’s framing, approx.) (Sam · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 33:58)
- Duration of the Meta AI support exploit before detection: 4 months (host’s framing) (Sam · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 25:25)
Caveats & warnings
- Sam’s bot-rebellion model is a thought experiment: he says it is not the business model the labs pitch or what they do, and not their current plan. (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 32:17)
- Dave’s claim that US frontier models are hobbled and that Kimi was needed rests on back-channel engineering talk and an ‘I think’ about K3. (Dave · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 15:00)
- Sam does not know whether the 80 billion dollar Chinese social engineering figure is true; he heard it in a meeting. (Sam · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 33:55)
- Dave doubts independent auditing with employee badges would be easy to coordinate or trust. (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 17:09)
- Sam says he is not a doomer who expects a paperclip-maximising AGI, though he expects chaos. (Sam · Anthropic Wants to Slow Down. (Sep 2026) ▶ 09:48)
- Host has a stake: Dave co-founded Sauron and pitches it to Sam (‘Invest in Sauron, Sam’); he says he doesn’t talk about it publicly. (Dave · Apple’s iPhone Sock, Waymo’s B (Nov 2025) ▶ 26:21)
How much to trust this
- Evidence: 18 episodes, 2 independent speakers (the four hosts and unidentified host speech count as one voice, the show; guests counted individually).
Blueprints that use this
Sources
- (01)
Is Google About to Kill Canva? │Claude Goes Rogue, Tesla vs. Waymo
15:55 into “Is Google About to Kill Canva? │Claude Goes Rogue, Tesla vs. Waymo”
- (02)
OpenClaw vs Meta vs OpenAI: The Personal Agent Wars Heat Up
44:18 into “OpenClaw vs Meta vs OpenAI: The Personal Agent Wars Heat Up”
- (03)
Stripe Is Trying To Buy PayPal For $150b | OpenAI Cyber Incident, Frontier Model Economics, Peptides
- (04)
Open Source AI Wins. Now Labs Are Running to Washington |Thinking Machines, Johnny Ive, Realtime API
- (05)
Founders Beware: Illiquid Valuations vs. Real‑World Taxes
6:27 into “Founders Beware: Illiquid Valuations vs. Real‑World Taxes”
- (06)
Elon Musk vs OpenAI Trial, Google Cloud Surge, Meta’s Blocked Acquisition, Anthropic Winning
- (07)
OpenAI's $852 Billion Valuation Makes Zero Sense (ClawCon Tokyo, Anthropic's Leak, Goodbye Internet)
- (08)
Normal People Don't Want AI Agents (And Silicon Valley Can't Accept It)
10:24 into “Normal People Don't Want AI Agents (And Silicon Valley Can't Accept It)”
- (09)
Paramount vs Netflix: Who Really Wins Warner Bros And What's Next?
40:46 into “Paramount vs Netflix: Who Really Wins Warner Bros And What's Next?”
- (10)
SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.
33:59 into “SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.”
- (11)
The Real Reason Elon Wants SpaceX and Tesla to Merge Now
3:12 into “The Real Reason Elon Wants SpaceX and Tesla to Merge Now”
- (12)
Jeff Dean Leaves Google, Airtable Sells, and the Bag-Securing Era
32:57 into “Jeff Dean Leaves Google, Airtable Sells, and the Bag-Securing Era”
- (13)
Google I/O Just Killed SEO? Agentic Search, Universal Cart, and the New Web
43:54 into “Google I/O Just Killed SEO? Agentic Search, Universal Cart, and the New Web”
- (14)
What Tim Cook’s “Executive Chairman” Move Really Means | Musk vs OpenAI, xAI and Cursor Rumors
- (15)
Why SpaceX's $1.75T IPO Makes No Financial Sense (And Doesn't Have To)
25:07 into “Why SpaceX's $1.75T IPO Makes No Financial Sense (And Doesn't Have To)”
- (16)
Anthropic Wants to Slow Down. Is It Because They're Losing?
9:19 into “Anthropic Wants to Slow Down. Is It Because They're Losing?”
- (17)
Silicon Valley Is Great at Tech and Terrible at Business
20:32 into “Silicon Valley Is Great at Tech and Terrible at Business”
- (18)
Apple’s iPhone Sock, Waymo's Bay Area Expansion, and AI Wearables
26:21 into “Apple’s iPhone Sock, Waymo's Bay Area Expansion, and AI Wearables”