OpenAI rehearses for a possible AI crisis

Welcome to The Signal from Superpower Daily with Maya and Theo Right now top executives at major AI labs are quietly sitting in boardrooms Yeah rehearsing for catastrophic AI driven disaster scenarios And the massive public backlash that would inevitably follow Exactly I mean imagine waking up tomorrow and your banking app is offline The traffic lights are dead The local water supply is contaminated It sounds like a movie pitch It really does but it isn't We are doing a deep dive today into how the industry is preparing for these exact vulnerabilities Along with a massive shift in tech hiring Right And a terrifying exploit uncovered by AI

in the crypto space Plus the sudden very real realization that AI compute actually costs money Real money Yeah Let's get right into our lead story Sounds good So we have a new report from Axios and it details some very intense tabletop exercises happening right now behind closed doors At the major AI labs Right Like OpenAI and Anthropic Right Executives are simulating what happens if an AI enabled cyber attack causes major tangible public harm They are gaming out attacks on essential infrastructure Like power grids going down Yeah Water supplies compromised Banking systems failing Wait let me just clarify for our listeners a tabletop exercise That is

basically a corporate war game right Exactly It's an intensive interactive war game for leadership You put the CEO the head of security and the communications team in a room All together Right And you hand them a hypothetical crisis scenario So hour one a vulnerability is discovered Okay Hour two it gets exploited Hour three the New York Times is calling for a comment Oh wow So it moves fast It does But the crucial detail here is their focus on the aftermath They aren't just planning for the technical fallout Like they're just worried about a server going down They are explicitly planning for the day after The

day after the attack Yes They are preparing for a public revolt I mean that makes sense because the companies realized they would face a dual threat Exactly They would have this frantic operational crisis on their hands trying to patch a vulnerability Right But simultaneously they would face an absolute crisis of trust That is the defining element of these rehearsals Think about it If an AI model is used to hack a water treatment facility the public won't just blame the hacker No they will blame the toolmaker Right They will blame the AI company The executives know this So they are preparing for the sweeping government response

that would inevitably follow Because a major incident like this would permanently shape how lawmakers regulate artificial intelligence Without a doubt That would change everything But wait I have to ask if we should actually be worried here Are they planning for this like airlines plan for plane crashes What do you mean Well airlines know crashes are statistically possible So they have standard responsible risk management protocols Right But the Axios report mentions a six to 12 month timeline for a major incident So does that whisper mean they can actively see the engine smoking right now That is a great question And that timeline is the most important

caveat you have to keep in mind here That six to 12 month estimate is an insider forecast So it's a rumor Basically Yeah it is a rumor from industry sources It is not a company backed timeline And it is certainly not hard evidence of an imminent threat Right That is a big distinction It is OpenAI explicitly stated they do not treat these scenarios as inevitable They view them as hypothetical possibilities Exactly And Anthropic actually declined to comment entirely on the report So it's not a countdown clock It's just an acknowledgment that the landscape is shifting Precisely And OpenAI's perspective actually offers a bit of a

counterweight to the panic How so Well they acknowledge the AI tools are changing the cyber threat landscape But that change is exactly why they are focused on getting capable tools into the hands of defenders Oh I see Yeah They believe AI will ultimately strengthen cybersecurity if it gets deployed correctly by the good guys They aren't just sitting around worrying about attackers That makes sense But it still signals a new era for these companies It definitely does I guess the question is are they facing a uniquely AI shaped threat Or are they just maturing Right Because these preparations might just resemble the standard stress tests that

banks and the Pentagon have used for decades And that is exactly what you should watch next We really need to see how much these preparations actually differ from established crisis rehearsals Because these are young fast growing labs Exactly They might just be adopting the standard corporate risk management practices of a mature industry They are finally growing up The definitive takeaway here is that AI companies are no longer just building technology They are actively preparing to manage the political and social fallout if their tools are weaponized against critical infrastructure Perfectly said All right Moving on to our first top story We are shifting from the potential

risks of AI to the rapidly changing landscape of the people actually building it Yes We are looking at the tech workforce specifically the data on software job postings Right Because it reveals a fascinating and honestly somewhat alarming divide between senior and junior roles It really does So U S software development job postings have actually risen up nearly 15 since late February 2025 Now that headline number is optimistic It kind of bucks the broader economic trend Yeah because overall job postings actually fell 7 in that same period Exactly But when you look closely at the Indeed hiring lab data the rebound is highly concentrated Right It's

tilted toward experienced workers Heavily tilted Senior roles drove 71 of the increase in software postings between May 2025 and May 2026 71 Yeah I mean that completely debunks the binary myth that AI is just replacing human coders Right It isn't a simple replacement No It's a fundamental shift in the nature of the work Companies are desperate for senior talent And to understand why that is you really have to look at how code is actually being written today OK So GitHub data shows that one in three pull requests now involve an AI agent Oh wait Just to clarify a pull request is basically a developer proposing

a change to a project's main code base right Exactly It's how new code gets added And last year AI agents were involved in fewer than one in 10 of those requests Wow So the volume of code being generated has just exploded It has And because companies are producing vastly more code they desperately need experienced senior developers to oversee it Right They need humans to manage the automated coding Yes They need experts to catch security vulnerabilities before they go live Think about an automated factory line OK Imagine a factory that suddenly installs these new machines Machines capable of stamping out parts 100 times faster than human

hands ever could I'm picturing it You don't need manual assembly workers anymore The machine does that But the role of the master mechanic becomes critical Oh I see You need experienced veterans to stand at the end of the line inspecting every single piece for micro fractures before they get shipped Right Because if a bad part ships it's a disaster Exactly The AI is the machine The junior developer was the manual assembly worker And the senior developer is the master mechanic Yes Their job is more important and vastly more stressful than ever before But wait It sounds great that companies are producing more code but if

junior devs cannot get a foot in the door aren't we just burning the bridge behind us That is the central tension right now Because you can't just spawn a master mechanic out of thin air They have to learn somehow Exactly And the Stanford 2026 AI Index highlights this perfectly They found that employment among software developers ages 22 to 25 actually fell 20 20 20 That is a staggering drop for early career developers It really is Now we do need to note a few limitations in this data Right That 15 rebound in job postings is relative Software postings are still 27 5 below their pre pandemic

levels Yeah the market hasn't fully recovered to those peak days And you also have to distinguish between job postings and actual employment That is a crucial point Postings measure advertised demand Employment measures who actually has a job Right And the AI titled roles only accounted for 37 of the posting increase So AI is a major factor but it isn't the sole driver of the rebound Companies still need traditional software engineering for a multitude of reasons Absolutely So here is what you should watch next Call it the talent pipeline time bomb Today's entry level talent becomes tomorrow's senior experts If entry level hiring collapses right now

where will employers find those experienced master mechanics a decade from now It is a massive problem looming on the horizon It really is And speaking of the critical need for senior oversight and security checks Yes Our next story shows exactly what happens when automated systems find deeply buried flaws that humans missed In other news we have this wild discovery in the cryptocurrency space It is genuinely terrifying It is So VariaLabs disclosed that its AI agent found a massive vulnerability in the XRP ledger Right This bug could theoretically create 18 trillion XRP tokens in a single transaction 18 trillion Trillion with a T It's a real

world demonstration of what AI agents are now capable of It really is So the bug sat in a piece of software called RippleD which is the core software running the XRP ledger OK And the vulnerability depended on chaining together two separate integer overflow defects OK An integer overflow That is a classic software problem right It is Think of an old mechanical car odometer OK The ones that max out at 99 999 miles Right You drive one more mile and it doesn't show 100 000 No it rolls completely over to zero Exactly Software memory works the exact same way It has physical limits If a calculation

results in a number larger than that fixed memory can hold it wraps around Right It can wrap back to zero or even into a negative number depending on the system design Wow OK So what did the AI actually do here The AI agent found two of these obscure overflow errors and linked them together Two of them Yeah The first error happened in the ledger's built in exchange Payment routed through several exchange offers could make the buyers running total overflow So what does that mean in practice It means the sellers got paid their full amounts but the buyer was charged almost nothing Oh Well that seems

like a massive problem on its own It does But on its own that first error should have been caught by the system's safeguards The ledger has a safeguard specifically called XRP Not Created Because XRP has a fixed supply Exactly Transactions are never supposed to create more tokens They actually destroy a tiny fee OK So the safeguard adds up all the balance changes and it checks to ensure the only net difference is that destroyed fee So the safeguard should have stopped the transaction when the buyer's total didn't match It should have But the AI found that the safeguard itself had the exact same storage limit flaw

Oh wow Yeah Its own mathematical ledger could overflow So when the first bug created a massive artificial supply increase The safeguard's math wrapped around It read the enormous increase as a tiny negative number Wait really Yes It essentially mistook the creation of trillions of tokens for an ordinary fee payment The ingenuity here is just terrifying I mean it isn't just finding a broken lock It's finding a broken lock and then realizing the alarm system connected to it has a loose wire you can short out simultaneously That is the perfect way to describe it Each defect would have stopped short on its own The AI realized

they could be chained together to defeat both checks That is wild So VariaLabs reported this on September 22nd and they received the maximum bug bounty of 250 000 Which they absolutely deserved And the speed of the response forced an emergency protocol that actually broke decentralized norms Right Because normally changes to the XRP ledger go through a strict amendment process Validators vote on changes over several weeks Yeah it's a foundational democratic process for the network Bypassing it is like declaring martial law in the crypto world But they had to They bypassed the usual validator voting process entirely Varia says this was the first time that process

had been bypassed in more than a decade Because the risk was just too high Right Now the most important caveat here is that RippleX investigated and they found absolutely no evidence that this flaw was ever exploited in the wild Right Varia built and demonstrated the exploit locally on a testing network And the specific build configuration of the software was also key right Yes The official release binary allowed the arithmetic to silently wrap around Okay But if the software had been configured differently it would have just crashed the node instead of minting the tokens So it was a very specific set of circumstances It was But

what you really need to watch next is how open source security adapts to this new reality Yeah because the developers had to distribute ready to run binaries before they published the actual source code Exactly They temporarily withheld the code And open source communities pride themselves on total transparency They do But AI is getting so good at finding vulnerabilities humans are having to abandon that transparency just to board up the windows in time The wild irony is that AI is forcing closed door behavior in open source systems It really is Next up The incredible capabilities of these AI tools are driving unprecedented demand Which brings us

to a major physical bottleneck over at Anthropic Yes Anthropic has paused its free year of the CLOD team plan along with its 1 000 API credit offer for startups They received hundreds of thousands of applications in just a few days Hundreds of thousands Yeah The surge completely overwhelmed the program's capacity So approval for the program no longer guarantees founders those two headline benefits No Unclaimed approved accounts are being put back under review And Anthropic stated clearly that some previously approved founders will simply lose out on the offer It really highlights the sheer physical reality of compute constraints People so often confuse AI with traditional software

as a service Right SAS has massive margins You can give away infinite Dropbox storage or Slack seats because the marginal cost is basically zero But AI isn't like that at all Not even close When a startup uses an API credit to query a model like CLOD it requires physical GPUs to process that request It consumes massive amounts of electricity It generates heat that requires water cooling It costs real money every single time a prompt is processed Every single time Even heavily funded Frontier Labs cannot absorb unlimited free API calls Giving away a year of a team plan and 1 000 in compute to hundreds of

thousands of startups I mean that is an astronomical expense It is And it signals a major pivot for the industry We are seeing a shift from blanket Silicon Valley growth tactics to a much more scrutinized ROI focused approach to developer acquisition It makes you wonder did they really just underestimate demand Or did someone in finance look at the projected Amazon Web Services bill and pull the fire alarm Honestly it's likely a mix of both The era of limitless free compute is ending But there is an important limitation to note here Right This does not shut down the broader CLOD for startups program Founders who already

redeemed their offers get to keep them Yeah And the startup stack which offers up to 45 000 in third party discounts that remains active There is also still a path for VC backed startups in Anthropic's network to get up to 100 000 in credits But that requires specific venture backing It isn't an open application process for everyone Right Watch how Anthropic reshapes the program post review Look at exactly what criteria they use to filter the backlog of applications Will they prioritize specific industries How they choose the winners will tell us a lot about their long term business strategy All right Let's move to our quick

reads section Sounds good First up Bill Gates is weighing in on the most immediate dangers of artificial intelligence In an interview with ABC News the Microsoft co founder warned that AI enabled cyber attacks are the most immediate threat we will see over the next year Right He specifically highlighted the danger of people using models that have had their safeguards deliberately removed And he pushed back hard against the idea of company self regulation He argued that society needs wider checks against AI assisted cyber attacks and bioterrorism The key distinction here is his timeline He didn't dismiss the sci fi risk of AI eventually escaping human control

But he placed that major existential impact years away Right His warning right now is entirely about deliberate human misuse in the near term Meanwhile type safe AI just raised an 870 million Series A That is a massive round It is This vows the company at 7 5 billion just weeks after launching their new model Jev And Andreessen Horowitz led the round The fascinating thing about Jev is that it doesn't generate paragraphs of text or lines of code Right It's built strictly for decisions inside software It returns probabilities in a typed format Standard text generation is autoregressive It predicts one word at a time requiring a

massive neural network pass for every single syllable But Jev outputs a calibrated decision in a single pass Which is huge Because it handles this narrower task it is computationally lightweight The investors claim Jev is 100 times faster than frontier models for classification tasks And massively cheaper It generated a trillion tokens in its first three days But there is a glaring mismatch in the public relations messaging here What do you mean Well type safe says a third of the Fortune 500 are using it While Andreessen Horowitz says 25 have integrated it Oh Those terms mean very different things in enterprise software They really do Finally cybersecurity

firm Sophos has reported that their new AI agents cut security response times from roughly 38 minutes down to just 89 seconds 89 seconds That is incredible It is These agents were built using OpenAI Daybreak They live inside Sophos Fusion resolving 52 of managed detection and response cases entirely end to end The agents operate on a plan execute review loop They don't just summarize alerts Right The AI plans an investigation executes checks against the active directory or firewall logs reviews the results and prepares a response strategy It does in seconds what takes humans dozens of minutes of clicking through dashboards The crucial limitation here is boundary

control The agents operate strictly within customer defined permissions Right The customer chooses if Sophos can only notify them collaborate with them or authorize a response Human oversight remains mandatory for any potentially destructive actions As it should All right We are moving to three takeaways from today First AI security is an escalating arms race From the tabletop crisis exercises to the Sophos defense agents and the XRP exploit Artificial intelligence is radically compressing the timeline for both attackers and defenders Second the tech labor market is polarizing The massive volume of AI generated code is hollowing out the entry level path It is simultaneously driving up the premium

for senior developers who can actually oversee and secure all that automated output Third compute reality is setting in Whether it's Anthropic capping startup credits or TypeSafe raising nearly a billion dollars for a strictly decision focused software model the industry is aggressively searching for sustainable economics One development to watch tomorrow Keep an eye on how open source communities like the XRP ledger handle the tension between code transparency and security Now that AI can reverse engineer vulnerabilities instantly will the emergency practice of withholding source code become the new normal Head over to superpoweredaily com for more on all these stories Thank you for listening We'll see you

tomorrow