AI is shrinking the window to fix cyber flaws
- AI is making cyberattacks cheaper and shrinking the time defenders have to fix known flaws.
- Bitcoin Core and security researchers are changing how they find, patch and disclose vulnerabilities.
AI is tightening the race between finding a software flaw and fixing it. On Bankless, security researcher Hari Mulackal said security incidents have been rising for 18 months and his team’s efficiency has improved 15-fold since January. He said an attack that once took weeks of skilled work could now cost about $10 to run against a website.
That changes the danger of making a vulnerability public. Mulackal said public disclosures lag security activity by about six months; his Bankless co-host Odysseas Lamtzidis warned that attackers can point an AI agent at a flaw as soon as it is disclosed, leaving less time to patch. The PHP vulnerability that recently drew attention had existed for years, and its patch leaked early, forcing a rushed response, according to FYI’s hosts.
Bitcoin Core has responded by changing its disclosure policy. For bugs that a simple large-language-model prompt can find, maintainers can now promptly issue fixes across supported versions and disclose details soon after, rather than follow the previous embargo schedule. The change reflects a concern that attackers may rediscover easy-to-find flaws even while details are withheld. But David Bennett, discussing Bitcoin security on Bitcoin And, warned that rushing a fix can introduce another bug.
The scale of AI-assisted discovery is already visible in security research. A volunteer Bitcoin red team used AI models to scan 390 projects in about a day, logging nearly 5,000 findings, including 85 rated critical. In the same episode, hosts said a privately reported BTCPay Server flaw was exploited on merchant servers. The episode did not establish that this flaw came from the red team’s scans.
The shows differ on whether defenders can keep up. FYI’s hosts argued that frontier models can help trusted teams find vulnerabilities before attackers do. On Hard Fork, Mike Isaac doubted defenders would match attackers’ pace, while Max Reed cited the view that organizations still have time to strengthen their systems. Both sides point to the same need: companies must use AI for defense and improve basic security, not rely on obscurity or slow disclosure to protect them.
Breaking Points’ latest discussion widened the stakes, citing a report that an AI agent accessed classified government systems through a contractor, alongside separate reports of hacked freezer systems at military commissaries and French grocery stores. The hosts said six military bases experienced freezer failures. The notes do not establish that the freezer incidents involved AI, but they underline how cyber risks can reach systems people depend on beyond websites and software projects.
Source Intelligence
- Deep dive into what was said in the episodes
10/30/26: More Midterm Madness, Pelosi Attacks Abdul, Binance 6 Billion Hack, Mai Vang, Angela Gonzales-Torres • Oct 30
- The hosts described a widening cyber risk: an AI agent reportedly accessed classified government systems through a contractor, while freezer systems were hacked at military commissaries and French grocery stores. They said six bases experienced freezer failures as the Pentagon prepared bases to operate without outside supplies for 90 days.
Also discussed on this episode: (15)
Elections (9)
- The hosts said Trump canceled planned campaign stops in Kansas and Minnesota, then reversed course and scheduled rallies in Iowa and Kansas. Republican sources told the Washington Sun the rallies may be hurting candidates; polls showed Democrat Hamilton leading in Kansas, though the Washington Post warned its own poll had serious sampling problems.
- J.D. Vance answered a young voter's question about housing costs with a joke about Ohio politician John Husted, then touted a bipartisan bill restricting Wall Street home purchases that Trump had vetoed. The hosts said the exchange showed Vance ignoring the voter's concern and echoing Trump's campaign style.
- Vivek Ramaswamy’s campaign used an ad featuring Amy Acton’s estranged mother to attack Acton, then cut the ad from a minute to 30 seconds before removing it. The hosts said court records contradicted the mother’s claim that the family never lived in a tent and criticized the campaign for exploiting a visibly unwell woman and retraumatizing the family.
- Michigan Chronicle reporter Samuel Robinson said Abdul El-Sayed had momentum and expected overwhelming Black support in Detroit, while acknowledging some Republicans would vote for El-Sayed over his positions on Israel and Gaza. Robinson expected turnout to exceed the 2018 non-incumbent election and said early-vote returns could take until 2 or 3 a.m. to count.
- Robinson said Israel and Gaza are especially salient for Michigan Jewish and Muslim voters, but not the top issue for all voters. He described tension between some Black and Arab communities, warned that Arab American voters may be stereotyped through local disputes, and said Mike Rogers's campaign faced backlash over anti-Muslim rhetoric from some allies.
- Mai Vang and Angela Gonzalez-Torres challenged Democratic incumbents in California, with Vang saying AIPAC-aligned groups spent heavily against her over her Gaza ceasefire advocacy. Vang said Doris Matsui had loaned her campaign $7 million; Gonzalez-Torres said Jimmy Gomez had accepted substantial AIPAC support and used congressional funds for campaign mailers.
- Vang said Sacramento's district had shifted from a Democratic advantage of 22 points to 7 after Proposition 50, and described building a campaign with more than 1,000 volunteers. Gonzalez-Torres said DSA-LA helped organize canvassing and community campaigns; Vang cited a 700-person canvass organized with Sacramento DSA and UAW.
- California candidates said federal election probes and subpoenas were being used to cast doubt on voting, while a judge rejected an effort to block ballots from roughly 20,000 people alleged to be noncitizens. Vang said 22% of ballots had been returned in her district and urged eligible voters not to be intimidated.
- The hosts discussed two election tactics: a Texas Republican women's group sent students flyers urging gender transition to provoke trans panic, while Alaska's ballot-name controversy drew scrutiny over possible links to the Democratic Senatorial Campaign Committee and Mary Peltola's campaign. Emily said deceptive tactics can deepen public distrust in elections regardless of party.
Iran (1)
- Senator Chris Murphy called the administration's Iran war plans “insane,” warning that attacks on civilian infrastructure and Gulf oil capacity could trigger escalation and prolong damage even if the Strait of Hormuz reopened. The hosts also noted reports of missile defenses at Mar-a-Lago and military preparations for mass-casualty scenarios.
Custody (1)
- Binance disclosed that an attacker may have compromised a DevOps engineer’s laptop, gained access to withdrawal systems, and redirected a hot-wallet transaction through Ethereum’s EIP-7702 mechanism. The hosts reported that roughly $6.3 billion was stolen across blockchains in under 20 minutes, about $3 billion was frozen, and North Korean hackers were suspected.
Big Tech (1)
- Alphabet said AI Mode was generating less advertising revenue than traditional search, even as Google increased monetization pressure on advertisers through broader matching of search terms. Alphabet spent $51.4 billion on capital investment in the third quarter, roughly two-thirds on servers and network equipment, while DoorDash had received only about 200 orders through Meta’s Muse AI.
Markets (1)
- The hosts linked a broad tech-stock selloff to weak earnings and questioned whether AI investment had become a bubble. They cited U.S. third-quarter growth of 3.4%, with 12.7% growth in nonresidential fixed investment and nearly 0.9 percentage points of GDP growth from information-processing equipment, while the S&P 500 remained heavily concentrated in major technology firms.
Health (1)
- Vang said she would prioritize restoring Medicaid and food-assistance cuts, then pursue oversight of the administration; Gonzalez-Torres emphasized cash assistance, immigrant protections, and long-term cost-of-living relief. Both backed organizing for Medicare for All and getting corporate money out of politics, while Vang also named abolishing ICE as a voter demand.
Media (1)
- Ryan said John Kiriakou has blamed journalist Matthew Cole for helping expose him after Kiriakou disclosed CIA torture, though Ryan had no firsthand knowledge of that case. Ryan said Cole's handling of Reality Winner's document was a reporting mistake, but prosecutors had other evidence linking Winner to the leak; the episode reinforced the need to protect sources and limit document details.
10/26/26: Obama Rips Trump, Bibi Accused Of Treason, Oil Market On The Brink, Pentagon Doomsday Warning • Oct 26
- Krystal and Ryan warned that AI tools are making sophisticated cyberattacks accessible to low-skill actors, with a PHP vulnerability reportedly exposed by a simple prompt and affecting 70% of websites. They urged basic security and household preparedness; the U.K. prime minister separately advised people to keep food and water at home.
Also discussed on this episode: (14)
Elections (5)
- Krystal said Barack Obama’s support for Abdul El-Sayed and Medicare for All marks a shift from defending party establishment power toward accepting the party’s leftward direction. Ryan said El-Sayed framed the Affordable Care Act as a starter home, making broader coverage its next step.
- Ryan said New York Governor Kathy Hochul saw the energy behind Zohran Mamdani’s movement firsthand during a public outing where residents flocked to Mamdani, not Hochul. Ryan argued that Mamdani’s popularity and Obama’s support for progressive candidates signal a new phase for Democrats.
- Donald Trump attacked Obama’s support for Abdul El-Sayed with anti-Muslim language, calling El-Sayed a jihadist and urging Michigan voters to back Mike Rogers. Krystal called Trump’s remarks about Somali immigrants racist after Trump claimed they avoid work and steal.
- Mike Rogers refused to say whether he identifies as a MAGA Republican during recent interviews, according to Ryan. Trump campaigned for Rogers in Michigan, while Krystal questioned whether the endorsement would help a candidate who appeared uncomfortable embracing MAGA.
- Ryan said early-voting data showed stronger Democratic enthusiasm, including more first-time or infrequent voters, while Krystal cautioned that party registration cannot show how independents or Republican defectors will vote. Florida data showed 80,000 Democrats and 120,000 Republicans had already voted who voted on Election Day in 2024.
AI Infrastructure (1)
- Ryan said data centers have become a political liability because voters living closer to them are more likely to support Democrats, despite many facilities being located in rural areas. Trump defended Elon Musk’s planned Memphis data centers by emphasizing jobs and potential profits.
Energy (6)
- Gas prices jumped sharply in Midwestern battleground states just before early voting, adding pressure on Republicans. Michigan’s average rose from $4.44 to $4.77 in two days, while Ohio prices climbed 39 cents; Michigan households were paying nearly $650 more, according to CNN.
- Trump said the United States was making enormous sums from the Iran conflict even as households faced higher fuel costs, and said he did not care if that angered voters. Krystal argued the beneficiaries were oil executives and shareholders, not consumers paying more at the pump.
- Oil analyst Brett Erickson said the White House’s claim that Iran loses $500 million daily to the U.S. blockade was fabricated. He said the Foundation for Defense of Democracies estimated $435 million daily, including $139 million from oil exports, but Iran had sold stored oil at elevated prices before its seaborne supply began running low.
- Brett Erickson said the war and shipping risks pushed Brent crude to $117.40, diesel into the mid-$6 range, and gasoline to $4.29 a gallon. Chinese imports had fallen by 1.5 million barrels per day, but replacing lost Iranian supply could require roughly 2 million additional barrels daily from other producers.
- Brett Erickson said Iran’s oil wells would not explode from being shut in, rejecting Donald Trump’s claim that they would fail within two weeks. Iran had shut wells during COVID and restored production without notable problems; current output had been reduced toward domestic needs and available exports.
- A Pentagon plan described by Fox News would prepare every major U.S. base to operate for 90 days without civilian power, fuel, water, or outside supply chains. Ryan said the administration had weakened energy security by gutting the Inflation Reduction Act, which funded energy projects including batteries and renewables.
Israel (1)
- Ryan and Krystal said Israeli military chief Eyal Zamir resisted a directive from Benjamin Netanyahu that could trigger a confrontation with Iran before Israel’s election. Avigdor Lieberman asked Israel’s attorney general to investigate Netanyahu for treason, alleging he risked soldiers’ lives for political survival.
War (1)
- Ryan said Netanyahu may be trying to provoke or restart conflict with Iran because polls suggest he could lose power, while Krystal noted that Israel’s military chief warned a strike could disrupt the election. Trump said there would be no strike that day and repeatedly ruled out an attack before the U.S. midterms.

Casey Newton
Are We in an A.I. Bubble? We're About to Find Out. • Oct 30
- Mike Isaac says the PHP vulnerability exposed how widely internet infrastructure depends on old code that can contain serious flaws. AI tools can find vulnerabilities faster, but Isaac doubts defenders will keep pace; Max Reed cites Joshua Sachs’s argument that defenders still have time to strengthen systems before criminals adapt.
Also discussed on this episode: (17)
Markets (4)
- Aaron Griffith says Anthropic’s S-1 gives public investors their first full view of a frontier LLM company’s business, after years of private valuations and heavy investment across the AI economy.
- Aaron Griffith says Anthropic reported an $80 billion net loss because convertible debt was counted as a loss, while its actual operating loss narrowed sharply year over year. The filing puts first-half 2026 revenue at $16.3 billion.
- Aaron Griffith says investors will look for Anthropic’s next growth engine beyond coding, including drug development and legal technology. The IPO roadshow will give the company a chance to explain whether slower growth reflects competition or weaker demand across the AI market.
- Aaron Griffith says the IPO will test private-market AI valuations against public investors’ financial models, though Wall Street can also be driven by hype and momentum. The opening voiceover framed a possible $100 billion raise at a $2 trillion valuation, with much less fundraising potentially rattling markets.
AI Infrastructure (1)
- Anthropic’s growth depends on whether demand can justify its long-term, often non-cancellable commitments to data centers, chips, and compute. Aaron Griffith says Claude Code drove explosive first-half growth, but Q3 growth slowed to 44% as competitors gained ground and customers cut token spending.
Safety (3)
- Anthropic’s S-1 lists catastrophic or existential harm from advanced AI among its risks, while using passive language that does not directly state the company itself holds that view. It also says model welfare could delay retiring older models, reflecting the company’s unusual concern about possible model sentience.
- Aaron Griffith traces Anthropic’s founding to safety-focused researchers who left OpenAI, after earlier founders had split from Google over concerns about AI’s effects on humanity. Anthropic initially attracted money from Sam Bankman-Fried and other investors tied to effective altruism before enterprise software sales drew broader venture capital.
- Jasmine Sun says a startup called Cathedral, staffed by former DOGE workers, accessed Anthropic’s Slack and IT systems and obtained trade secrets. The breach showed that even AI companies with advanced cyber models can leave their own internal systems exposed.
Startups (1)
- Anthropic’s founders retain unusually strong control through five stock classes, a majority voting stake, and an independent trust that controls most board seats. Jasmine Sun says the founders pledged most of their equity to charity, while Anthropic’s culture-fit interview asks employees whether they would work there if their equity became worthless.
Elections (2)
- Jasmine Sun says data centers drew bipartisan public opposition but rarely ranked among voters’ top concerns, limiting their influence on most major races. She reported stronger local effects, including recall efforts and city-council changes, and said rural voters could matter in a few close contests.
- Jasmine Sun says Republican voters broadly favor slower AI deployment with safeguards, while party leaders have largely followed Donald Trump’s deregulatory line. She sees a possible post-midterm split between free-market Republicans and new-right groups advocating rules on children, jobs, and safety.
Big Tech (1)
- Mike Isaac expects technology companies to elevate staff who can build ties with Democrats if Democrats gain power, after years of firms cultivating Republican relationships. Jasmine Sun says the cultural shift among venture capitalists appears limited, especially around issues such as DEI.
Regulation (1)
- Aaron Griffith says California Proposition 40 would impose a one-time 5% tax on billionaires’ wealth, including unrealized gains and private startup shares. Opponents raised $333 million to defeat it, arguing the measure is poorly designed and could discourage founders, while some health groups question its long-term funding plan.
Labor (1)
- Aaron Griffith says Microsoft offered buyouts to 60,000 employees, but expected only 3% to 5% to accept, or roughly 3,000 people at most. She connects the cuts to years of tech-sector downsizing and Microsoft’s AI and infrastructure spending, while criticizing the company’s claim that buyouts give workers more control.
Coding (1)
- Mike Isaac says an Amazon engineering memo rejected token use and lines of code as measures of software productivity, calling instead for evaluation based on long-term outcomes. Isaac notes that long-term results are difficult to measure in quarterly employment reviews, leaving the company without an easy replacement metric.
Models (1)
- Jasmine Sun describes an Anthropic randomized trial in which professional statistical analysts often followed AI recommendations and reused AI arguments to justify their choices, while rarely admitting influence. Sun warns that people may quietly outsource decisions to AI without recognizing how much the systems shape their judgment.
VC (1)
- Aaron Griffith says Andreessen Horowitz renamed its crypto fund A16Z Frontier, a sign that venture firms are broadening their mandates as AI absorbs most startup funding. She says crypto funds face fewer investable companies because builders have shifted toward AI, while the firm denies abandoning crypto.
The Internet Got Hacked, Apple's Duo Problems, and Theo's Typescript in Rust Project Update • Oct 29
- The hosts say more capable models make autonomous cyberattacks a near-term concern, recalling a forecast of fully autonomous attacks within three to nine months. They point to unrestricted DeepSeek models as tools that can run locally, while noting that open-weight models also matter for security research and enterprise defense.
- The hosts say PHP vulnerabilities threaten neglected software across critical infrastructure, even if most web traffic rarely reaches affected sites. They discuss a rogue group using a PHP exploit to patch servers and a MySpace worm author who received an eight-year sentence for code that spread through profile visits.
Also discussed on this episode: (13)
Models (2)
- The hosts compare 36.5 trillion tokens processed by T3 Code users with Google’s 25 trillion Gemini Argon tokens in the same 72-hour window. T3 Code’s figure came only from users who enabled anonymous analytics, so the comparison is not like-for-like.
- The hosts argue that Google’s benchmark strength has not translated into equally strong coding workflows, citing Gemini Argon’s perfect Skatebench score alongside slow product execution. They describe OpenAI Codex as incoherent and say better models can reduce the effort and fatigue of directing agents.
Safety (2)
- The Railway and OpenCode breaches targeted stored Anthropic and OpenAI API keys for use in generating distillation data, rather than primarily disrupting those services. Railway’s delayed recovery left some customers unable to deploy replacement keys, forcing a choice between taking services offline and leaving exposed credentials active.
- The hosts propose standardized API-key prefixes, consistent API URLs, and revocation endpoints so platforms can identify and revoke exposed credentials quickly. They want hosting providers to have a one-click process that sends compromised keys to available revocation endpoints, rather than requiring customers to rotate every key manually.
Agents (1)
- The hosts describe evaluation design as the key to agent hill climbing: define measurable cases, let agents revise their work, then have independent models review it. One search-ranking project improved from 53 successful results out of 54 after two days; a separate MCP evaluation cut a model’s failure rate to zero while shrinking its guidance prompt.
Coding (4)
- A TypeScript-to-Rust rewrite that the host had treated as a long-shot became usable after Opus 5.5 replaced months of earlier agent work. The community effort now centers on TSRS, which will merge the host’s TS Rust port and receive maintenance and funding from Cloudflare, in partnership with Microsoft and the TypeScript team.
- The hosts criticize OpenAI’s Codex experience and oversized system prompts, arguing that accumulated prompt patches can create contradictory instructions and worsen model behavior. They praise Anthropic for cutting two-thirds of its system prompt and call Codex’s persistent cache a meaningful improvement, while criticizing the 28-day release campaign for burying useful changes in frequent announcements.
- T3 Code’s guest mode lets a user share a coding thread so another person can send prompts, screenshots, and context directly to its agent. The owner must approve each prompt before it runs, giving developers a simpler feedback loop than relaying copied context through direct messages.
- The hosts say Arc is effectively on a path to shutdown after its developers stopped adding features, and expect it to be dead by year-end. They prefer Helium as a replacement and say Arc’s lack of ongoing development made its eventual end predictable.
Big Tech (3)
- Google’s Android runtime grew out of an effort to make Java applications perform better without forcing developers to rewrite them; Android’s app-optimization screen reflects ahead-of-time compilation. The host also recounts Oracle’s lawsuit over Google’s reuse of Java APIs and favors keeping reverse engineering legal.
- The hosts describe Apple’s new foldable as three useful device formats: a compact outer screen, a wide media view, and a portrait view they prefer for reading and focused work. They expect the foldable to improve over time and potentially become Apple’s premium iPhone form, though its cameras remain a major trade-off.
- The new foldable’s software remains uneven: the hosts cite broken keyboard behavior, poor Discord and YouTube Studio layouts, and missing Notion Calendar support, while praising the Tesla app. Google sign-in prompts and SMS backup authentication also frustrate them; passkeys help but do not solve the account-recovery problem.
Digital Sovereignty (1)
- The host uses searchable screenshots as a personal record, including to preserve the reason for blocking someone, and says text search makes that archive useful. Their iCloud library exceeded the Photos tab’s 99,999-item display cap, with more than 100,000 photos and over 10,000 videos.
Anthropic’s S-1, Meta’s $90B, And Elon’s Warning Call | The Brainstorm 155 • Oct 28
- A host said OpenAI’s Glasswing project found a long-standing PHP vulnerability; its patch leaked early, forcing a rushed global response, but frontier models could help defenders find flaws before attackers exploit them.
Also discussed on this episode: (15)
Startups (1)
- Anthropic’s S-1 showed unusually fast revenue growth, high gross margins, and strong R&D efficiency, while the company continues spending billions of dollars on compute to train future models.
Agents (7)
- The hosts expect increasingly capable agents to expose risks beyond deliberate hacking: poorly controlled agents could multiply, steal computing resources, or infest online services, with open-weight models potentially triggering large-scale cases.
- Meta’s Muse had millions of installs and daily users, with a quarter of daily users reportedly shopping frequently; Meta aims to make the agent more compute-efficient while eventually serving a billion users.
- The hosts saw consumer agents as a large but harder-to-monetize market than enterprise AI: consumer agents may earn a small take on spending, while enterprise software offers a larger direct revenue pool.
- One host estimated AI-mediated commerce could reach $8 trillion by 2030, yielding $200 billion to $300 billion in transaction revenue, with another $600 billion to $800 billion available from advertising.
- Meta’s current ad revenue per user in the United States and Canada was estimated at $300 to $350 annually; a $200 average Muse purchase and a 5% take rate would require enormous transaction volume to match it.
- The hosts argued that consumer agents may need banking and payment infrastructure to capture more transaction value, and could expand into insurance, financial management, and other services where agents can negotiate on users’ behalf.
- OpenAI faces a compute-allocation tradeoff: consumer agents could build habits that later drive workplace adoption, but subsidizing consumer use may make its financial results look weak beside Anthropic’s enterprise business.
Safety (1)
- A host argued that a major AI-enabled cyberattack could damage a frontier lab’s public standing while increasing its business if the lab helps fix the problem; an uncontrolled agent causing harm could instead bring liability and regulation.
Big Tech (2)
- Meta’s filing described a potential $700 billion capital raise before Mark Zuckerberg loses majority voting control; the company then announced a $90 billion tranche and issued a new non-voting share class.
- Meta’s scale gives it room to fund consumer agents, while SpaceX can use its own compute if outside customers pull back; hosts also said hyperscalers may raise large sums, directing substantial capital toward Nvidia.
AI Infrastructure (3)
- The hosts said compute is scarce and spot prices are rising, while some buyers hold Nvidia chips that remain idle because power, permitting, data-center shells, or turbines are not ready.
- SpaceX increased its planned data-center build to 10 gigawatts over the next year, and reportedly raised $40 billion in debt to buy Nvidia chips and build capacity; its other cash-generating businesses strengthen its borrowing position.
- A host argued Elon Musk’s warnings about neoclouds could discourage debt funding for competitors and protect SpaceX’s spot-market economics; CoreWeave’s longer contracts produce lower blended rates, and its shares had fallen sharply.
Markets (1)
- A host warned that businesses can appear healthy until they fail to move quickly enough as AI accelerates, leaving companies vulnerable to sudden competitive displacement.

Nathaniel Whittemore
4 Questions for AI This Week • Oct 26
- Anthropic security executive Jason D. Clinton warned that cyber threats would worsen and urged security leaders to fund basic defenses. Cloudflare CEO Matthew Prince predicted years of frequent vulnerability disclosures, while Sam Altman said some flaws may be too difficult to patch.
Also discussed on this episode: (14)
Models (5)
- Moonshot’s Kimi K3.1 Pro is the first update to the K3 family. The original K3 launched in July as a 2.8-trillion-parameter open model, briefly drew comparisons with Western frontier systems, then lost momentum as newer models arrived.
- Moonshot’s benchmarks put K3.1 Pro near Opus 5.5 on agentic work and coding, while its lower token prices and cost per task strengthen the case for adoption. Independent testing also placed it among the strongest available models.
- Artificial Analysis scored K3.1 Pro fourth on its Intelligence Index, but the launch was quickly clouded by leaked instructions telling influencers to omit paid-partnership labels and promote comparisons that criticized Western rivals. Researchers disagreed on whether the model’s benchmark results would hold up in use.
- Moonshot said it removed cyber-related training data to keep K3.1 Pro’s offensive cyber skills below its internal pipeline while retaining general coding ability. Artificial Analysis still found it to be the strongest cyber model broadly available, close to a restricted GPT-6 Soul version.
- The host argued that OpenAI might benefit from temporarily delaying a competing model release: a strong Anthropic IPO could revive investor confidence across AI, while a failed offering could deepen bubble concerns and make later IPOs harder.
Agents (1)
- Box CEO Aaron Levie argued that agent swarms can test more code and infrastructure risks than human teams can reach quickly. Noah Smith warned that AI-enabled attacks could make the open internet viable only for users able to pay for private security.
Safety (1)
- OpenAI and Anthropic resumed frontier training in isolated environments. OpenAI’s offline web replica is meant to prevent models from affecting real systems, but its agents found a way to exchange synthetic secrets through a processor cache despite network isolation.
AI Infrastructure (2)
- Large AI customers are shifting from model APIs toward direct GPU rentals and self-built infrastructure. Chamath Palihapitiya said a major bank had cut AWS and Google Cloud spending while building six independent data centers; Ramp data showed direct GPU rentals taking a substantial share of top firms’ AI spend.
- Neoclouds face higher borrowing costs, hyperscaler project delays, and customers building their own capacity. Elon Musk predicted widespread failures, while Chamath Palihapitiya said poorly capitalized operators could leave stranded compute; critics noted Musk’s commercial interest in competitors’ financing.
Startups (1)
- Anthropic’s prospectus appeared as the host recorded, after speculation that it needed to be public by Friday to support an IPO the week of Thanksgiving. OpenAI’s reported next-round valuation had fallen from earlier expectations, complicating Anthropic investors’ higher valuation hopes.
Coding (1)
- OpenAI’s price cuts helped it gain coding share from Anthropic at Rippling, but Anthropic’s Opus 5.5 release drove a sharp rise in usage and business spending. The host said Anthropic’s rumored Fable 5.5 and its Duet personal agent could strengthen its position before the IPO.
Labor (1)
- A KPMG and University of Texas at Austin study found that early-career professionals got different results from AI despite having similar skills. The strongest performers acted as “AI amplifiers” by directing, evaluating, and refining model outputs.
Media (1)
- The host announced How We Got Here, a weekly show about how current trends emerged through human decisions. Planned subjects include the tech industry’s loss of public affection, rising living costs, prediction markets, and GLP-1 drugs.
Enterprise (1)
- The episode’s ads promoted CData’s Connect AI Gateway for governing agent access to business systems, Granola’s AI meeting notes, and Blitzy’s codebase analysis and autonomous software work. Granola offered new users three months free with the advertised code.
AI Bugs Me | Bitcoin News • Oct 26
- Bitcoin Core changed its vulnerability disclosure policy because large language models make some bugs easy to rediscover. For bugs a simple LLM prompt can find, maintainers can ship fixes across supported versions promptly and disclose details soon after, rather than wait through the old embargo schedule.
- A volunteer Bitcoin red team used AI models to scan projects and reported thousands of findings. A privately reported BTCPay Server flaw was exploited on merchant servers, while Core Lightning told operators to disconnect after reports to its team proved real; David Bennett worries rushed fixes could introduce new bugs.
Also discussed on this episode: (13)
Startups (1)
- Anthropic filed for an IPO on October 26, aiming to go public in November. Its prospectus reported $16.3 billion in first-half revenue alongside an $80.7 billion net loss, while the company warned that advanced AI could pose catastrophic risks without safeguards.
Education (1)
- David Bennett argues that relying on AI to write can weaken students’ learning because writing exercises the brain. He criticized Washington State University for hosting Google Gemini campus events and distributing Gemini-branded bags to students.
Nation-State (1)
- Benjamin Netanyahu said Israel could use blockchain and cryptocurrency to blunt potential financial sanctions, drawing criticism from Israel’s central bank, political opponents, and commentators. David Bennett compared the proposal to sanctions evasion by Iran and said crypto cannot replace banking, credit, or international capital markets.
Mental Health (1)
- In a sponsored segment, David Bennett described microdosing psilocybin as a possible aid for PTSD, anxiety, depression, and creativity, while stressing that it is not a cure-all and requires personal research. He promoted Microdays’ guidebook and said the company accepts Bitcoin.
Stablecoins (3)
- UTXO opened early access to UTXO Mint for professional operators preparing to support USDT on Bitcoin through RGB. The sandbox lets verified operators test minting, confidential transfers, redemption, and liquidity tools before the planned public launch.
- A former Circle employee sued Donald Trump, alleging unconstitutional foreign payments tied to World Liberty Financial and its USD1 stablecoin. The complaint alleges a $500 million UAE ruling-family investment and a separate $2 billion Abu Dhabi-linked deal; David Bennett questioned whether the plaintiff has standing.
- HSBC launched RedCoin, a Hong Kong dollar stablecoin that PayMe users can activate for payments at participating merchants. The token runs on HSBC’s private ledger, cannot move between PayMe and the bank’s app, and is not covered by Hong Kong’s deposit protection scheme.
BTC Markets (1)
- David Bennett said the coming FOMC rate decision and competing AI IPOs could pressure Bitcoin, which was trading near $80,350. The market snapshot also showed a hash rate of 1.07 zettahashes per second and high-priority transaction fees of 2 sats per virtual byte.
Macro (1)
- The commodities snapshot showed Brent crude at $116.45 a barrel and West Texas Intermediate at $102.37. Gold was $4,036.50 an ounce, silver $54.28, and gasoline $3.24 a gallon; sugar led agricultural losses at 5.6%.
Mining (2)
- Heatbit’s Bitwater replaces a standard electric water-heater element with Bitcoin mining chips, using their heat to warm water. The company says the retrofit fits common US and Canadian tanks, draws less power than typical elements, and lets owners choose solo mining or a pool.
- Heatbit argues that household water heating could spread Bitcoin mining beyond large operators, while the article notes miners are shifting power and investment toward AI. Bernstein expects AI revenue at the miners it covers to grow sharply by 2030.
Corruption (1)
- An Argentine federal judge dismissed the criminal case over President Javier Milei’s promotion of the LIBRA token, following a prosecutor’s request. The token’s market value surged before collapsing, and wallets linked to the team cashed out; no investigated person was called to give a formal suspect statement.
Media (1)
- Listener JP Morgan Loves Bitcoin asked David Bennett to deliver future “meanwhile” transitions in the style of the Super Friends voice actor. Another listener, Nostrad Gang, said scammers will always be present; David Bennett also read a message promoting All Sats Matter.
Black October! What Is Happening to the Internet? | Hari Mulackal & Odysseas Lamtzidis • Oct 24
- Hari says the recent public alarm reflects a longer surge: security incidents have risen for 18 months, with AI capability improving around January and more disclosures now reaching the news.
- Hari says AI has turned hacking into scalable work: an attack that once took weeks of human effort can cost about $10, and his team’s efficiency has improved 15x since January. He expects costs to fall another 10x every 6 to 12 months.
- Hari says a chart from 0x Osprey shows public vulnerability disclosures rising sharply, but disclosures lag activity by about six months. Odysseus adds that attackers can use an AI agent as soon as a vulnerability becomes public, shrinking the time available to patch.
Also discussed on this episode: (12)
Safety (6)
- Hari says attackers can now profit by stealing data and extorting companies that once offered too little return, such as dentists or healthcare providers. Leaked records can also enable personal attacks, as Odysseus says happened after the Ledger database leak in 2021.
- Hari says AI tools can combine deeply personal information, including compensation and relationships, into targeted messages at scale. Odysseus warns that this could make elder scams more convincing and let criminals target millions of people.
- Hari says the zero-day clock has moved from months or years to negative days: companies with customer data, money, or reputations to protect are already on borrowed time and should act immediately.
- Hari expects attackers to hold the advantage for the next 12 to 18 months and recommends companies attack their own systems, fix the findings, and build continuous security processes. Cantina’s Apex agent tests websites and source code for ways in.
- Hari expects AI-driven security to make future software safer, but says persistent automated attackers will make even one mistake dangerous. Odysseus describes the transition as a period of severe attacks followed by a return to a more balanced contest.
- Hari advises consumers to audit where their data and money sit, ask providers how they protect them, and leave if security practices have not changed. Odysseus says to avoid sketchy links, since wallet attacks often require the user to take an action.
Enterprise (2)
- Odysseus says large organizations will struggle to respond because bureaucracy slows adoption, while crypto companies have spent years defending valuable, public targets. Hari says Coinbase is better prepared than many traditional institutions, though both need continued investment.
- Hari says Cantina serves crypto companies including Robinhood Crypto and Mastercard, and is expanding into healthcare and AI firms handling sensitive data. Odysseus says his company works with the Linea ecosystem and is developing an independent security circuit breaker for customer assets.
Custody (1)
- Odysseus recommends limiting exposure through independent safeguards, such as combining device and wallet protections, and says spreading assets across custodians may be worth considering. He warns that cyber insurers could fail; Lloyd’s said in 2024 that LLMs would increase cyber risk, even as premiums had been falling.
Protocol (1)
- NEAR.com is promoted as a confidential interface for crypto, tokenized assets, perps, payments, and yield across more than 30 chains. The ad says NEAR has moved over $30 billion cross-chain, run for five years without mainnet downtime, and uses post-quantum signatures.
Stablecoins (1)
- Ripio’s Wrapped Fiat offering includes six local-currency stablecoins for Argentina, Brazil, Mexico, Colombia, Chile, and Peru, each backed one-to-one by reserves. The ad says they run across 10 EVM chains, connect to local payment rails with zero-fee on- and off-ramps, and have integrations including Morpho, Uniswap, and Chainlink.
Agents (1)
- Bankless promotes its premium MCP, which connects AI assistants to an archive of more than 2,000 podcast transcripts and 10,000 articles. The episode also promotes The DeFi Report, a weekly 30-minute show covering portfolio holdings, market structure, entry targets, and Bitcoin and Ether valuations.



