On the morning of October 30, the crypto market woke up to yet another theft of historic proportions. This time, Binance was the target.
According to the preliminary investigation, the attack may have begun with the compromise of a DevOps engineer’s laptop. From there, the attacker gained access to a service container and submitted unauthorized withdrawal transactions.
Binance estimated the incident’s total impact at approximately $6.3 billion. Around $3 billion in related assets has since been frozen through joint efforts with industry partners.
Major exchange security incidents have become uncomfortably frequent over the past year.
Bybit lost approximately $1.46 billion last year. In September, Bitget suffered roughly $387.5 million in unauthorized transfers. Now Binance has become embroiled in another major incident.
Viewed together, these events reveal a clear trend: attackers are increasingly avoiding direct attacks on cryptography and private keys. Instead, they are targeting the entire chain of trust—from employee devices and third-party security products to software supply chains, service containers, and transaction-authorization workflows.
A single crack anywhere along that chain can compromise the entire defense.
The rapid development of AI is magnifying this risk.
On October 22, Ethereum co-founder Vitalik Buterin offered a new assessment: cybersecurity’s long-term endgame may ultimately favor defenders, but the transition to that future will be extraordinarily difficult.
1. From Bybit to PHP2Shell: Attackers Are Compromising the Entire Chain of Trust
Let us begin by reviewing these incidents.
Their attack paths differed, but they had one thing in common: the attackers did not simply crack private keys or compromise the underlying blockchain protocol. Instead, they found a component that the system implicitly trusted.
In the Bybit incident, attackers used social engineering to compromise a Safe developer’s device, obtained relevant access credentials, and subsequently gained control of front-end resources within Safe’s wallet infrastructure.
In the Bitget incident, attackers exploited a vulnerability in a third-party security product to obtain highly privileged internal credentials. They then bypassed existing risk controls with forged withdrawal instructions.
In the latest Binance incident, the attack path may once again have begun with a compromised DevOps engineer’s laptop.
The three incidents involved third-party wallet infrastructure, a third-party security product, and an internal development and operations environment, respectively.
This helps explain why cold wallets and multisig mechanisms are not universal solutions. The attackers exploited trusted components capable of influencing the final execution of transactions but lacking sufficient independent verification.
And this crisis in the chain of trust is not limited to crypto exchanges.
One week before the Binance incident, Binance founder CZ specifically warned users to exercise extreme caution when visiting any website over the coming days. At the time, two critical vulnerabilities in the PHP core engine—CVE-2026-109618 and CVE-2026-109619, collectively known as PHP2Shell—were causing widespread alarm.
Both received a CVSS v4.0 score of 9.2.
PHP is an open-source server-side language used by roughly 70% of websites whose server-side programming language can be identified. Major platforms such as WordPress, Magento, and Drupal are built on it.
The two vulnerabilities, which had gone undetected for more than a decade, involved flaws in memory management and the resetting of global state. Under certain conditions, an unauthenticated attacker could send a specially crafted web request and achieve remote code execution before the website’s application code even began running, potentially taking control of the server, installing a backdoor, or stealing user credentials.
Scans by cybersecurity organizations including Palo Alto Networks and Censys suggested that between 22.6 million and 36.7 million internet-exposed instances could be at risk.
Even more strikingly, security researchers disclosed that AI tools had played an important role in uncovering these deep, long-hidden flaws in PHP’s memory and state handling.
In other words, AI has significantly lowered the barrier to turning obscure, low-level logic flaws into usable attack techniques.
We have previously discussed three fundamental ways in which AI is transforming cyberattacks. Further reading: “When Hackers Scale Up with AI: The Next Level of Web3’s Security Arms Race.”
- Industrialized zero-day discovery: Finding deep memory vulnerabilities in a core engine such as PHP or in device firmware once required elite security researchers to spend months on manual code review and fuzzing. AI can now scan global open-source codebases around the clock at far lower cost, potentially turning bugs that have remained hidden for a decade into exploitable weapons at scale.
- A drastically compressed patch window: Once vulnerability details leak or a patch is committed, malicious actors can use AI to reverse-engineer the fix, generate exploit variants, and automate internet-wide scanning and injection within minutes. Traditional operations teams that schedule upgrades over days or weeks may find themselves facing AI-powered attack scripts operating at machine speed.
- Software supply chains and DevOps as primary entry points: Modern Web3 and financial systems do not exist in isolation. Front-end interfaces, backend containers, third-party NPM and PyPI dependencies, and engineers’ local development environments form a vast and vulnerable open-source supply chain.
The reported compromise of a Binance DevOps environment and the front-end supply-chain attack behind the Bybit incident illustrate why the crypto industry is among the first to experience what could become a “pandemic-scale global wave of AI-enabled hacking.” Onchain attacks offer attackers an unusually direct path to monetization.
Of course, not every attack described above was launched or executed directly by AI. But AI is amplifying the capabilities of both attackers and defenders. Under today’s software engineering model, that amplification may initially intensify the structural asymmetry between the two sides.
2. Vitalik: The Endgame May Favor Defenders, but the Transition Will Be Brutal
Amid the industry discussion sparked by AI-assisted discovery of vulnerabilities such as PHP2Shell, Vitalik Buterin offered a measured assessment on October 22: cybersecurity may ultimately reach an end state that favors defenders, but the transition will be extraordinarily difficult.
Why?
Because the current software engineering model contains an inherent asymmetry between attackers and defenders.
A large software system may contain millions of lines of legacy code, hundreds of external dependencies, and countless complex permission and execution paths.
Defenders must secure every critical component. Attackers need to find only one undiscovered weakness.
Vitalik emphasized that simply relying on security researchers to find vulnerabilities in released software before hackers do cannot create a true defensive advantage. If both sides use increasingly powerful AI, the contest remains one of who can discover vulnerabilities faster.
What needs to change is how software itself is built.
He highlighted two core directions: formal verification and minimizing the Trusted Computing Base, or TCB.
Formal Verification: Replacing Trial-and-Error Testing with Mathematical Proof
Traditional software testing can show that no bugs were found in the scenarios tested. It cannot prove that a program contains no bugs under every possible input and state transition.
Formal verification takes a different approach. It uses rigorous mathematical logic to prove that software satisfies a defined security specification across all modeled inputs and state transitions.
This can help rule out classes of failures such as memory corruption, stale state—like the flaws involved in PHP2Shell—and logical reentrancy.
Historically, formal verification was prohibitively expensive. The mathematical specifications and proofs could require several times as much work as the underlying application code, making the technique practical mainly for aerospace systems, foundational cryptographic protocols, and other exceptionally critical software.
In the AI era, however, AI can do more than discover vulnerabilities. It can also become a powerful assistant for writing and checking formal proofs.
Vitalik therefore expects AI-assisted formal verification to become a central model for future software development. AI could eventually help generate end-to-end, machine-checkable security proofs for systems ranging from encrypted communication protocols such as the Signal Protocol and TLS to ZK-EVMs and application-layer smart contracts before their code is released.
When critical security properties have been mathematically proven, more powerful offensive AI cannot simply discover a violation that the proof has already ruled out.
Minimizing the TCB: Reducing the Attack Surface to Its Smallest Possible Core
A system’s Trusted Computing Base is the complete set of hardware and software components that must behave correctly for the system to remain secure.
The Binance incident and the compromise of Bybit’s multisig transaction process demonstrate why assets can still be stolen even when the underlying private-key cryptography remains intact: the TCB has become too large.
Container clusters, complex DevOps permission pipelines, developers’ laptop operating systems, and JavaScript libraries used to render transaction interfaces can all become part of a trust chain where a single compromise may ultimately lead to asset loss.
Minimizing the TCB means reducing the number of trusted components and isolating critical functions.
Functions directly responsible for private-key protection, signature verification, and critical authorization should be confined to a small, clearly defined security environment that can withstand rigorous auditing and verification.
More complex functionality should operate at the perimeter with limited permissions.
Even if a browser, plugin, front-end service, or AI agent is compromised, it should not automatically gain unrestricted authority over user assets.
From this perspective, the same principle also helps explain the value of non-custodial wallets such as imToken.
When users rely on a centralized exchange, their assets pass through platform accounts, custodial wallets, and internal withdrawal systems. With imToken, users hold their own private keys, manage onchain assets directly, and initiate transactions themselves, removing the centralized custodian from the trust chain.
A shorter trust chain means less exposure to platform-level risks, including exchange withdrawal suspensions and account-balance freezes.
Users must, of course, take responsibility for protecting their private keys and devices. But control over their assets no longer depends entirely on the internal operations of a centralized platform.
3. From Technological Acceleration to Defensive Acceleration: Why d/acc Matters
Formal verification and TCB minimization address the question of how technology can be made safer.
Vitalik’s long-standing advocacy of d/acc addresses a broader question:
As AI and other frontier technologies rapidly reshape the world, which forms of technological development should society prioritize?
d/acc stands for Defensive Accelerationism. Vitalik first presented the idea systematically in 2023 and expanded on it in his 2025 essay, “d/acc: one year later.”
The concept does not oppose technological progress. Instead, it advocates selective acceleration: society should prioritize technologies that strengthen defensive capabilities, distribute power, and reduce systemic risk.
Where e/acc emphasizes broad acceleration, d/acc asks whether a technology primarily strengthens offensive power or improves society’s ability to defend itself. Further reading: “When Web3 Meets d/acc: What Can Crypto Do in an Age of Accelerating Technology?”
Within the d/acc framework, defense spans four dimensions:
- Cyber defense: Using cryptography, formal verification, software isolation, and vulnerability detection to reduce the success rate of cyberattacks and protect digital assets and data.
- Biological defense: Improving society’s ability to respond to biological threats through pathogen detection, air monitoring, vaccine development, and public health technologies.
- Physical defense: Reducing the vulnerability of critical real-world infrastructure through decentralized energy, resilient communications networks, and hardened infrastructure.
- Information defense: Countering disinformation, deepfakes, and large-scale information manipulation through privacy protection, source verification, and decentralized coordination mechanisms.
These fields may appear very different, but they share a common objective: as technology increases the capabilities of individuals and organizations, they should also gain a corresponding ability to protect themselves. More powerful technologies should not be concentrated exclusively in the hands of a small number of resource-rich actors.
During the discussion surrounding the PHP vulnerabilities, Vitalik reiterated a forceful position consistent with the d/acc framework.
When frontier AI creates asymmetric risks, society should take a cautious approach to the uncontrolled development of high-risk offensive AI. At the same time, more than 1% of GDP should be systematically invested in cybersecurity, biosecurity, and hardening the physical world.
Why is investment on this scale necessary?
Because markets can quickly price and commercialize the benefits of technological progress, while the value of security infrastructure is frequently underestimated.
Companies are willing to spend enormous amounts on more powerful AI models, faster computing, and more efficient software-development tools. Yet they may be far less willing to provide comparable support for long-term maintenance of foundational open-source software, formal verification, and security audits.
Unless this misallocation of resources is actively corrected, the digital world will continue to grow taller on top of an increasingly fragile open-source foundation.
Final Thoughts
From PHP2Shell's exposure of systemic risks in global internet infrastructure to the major security incidents involving Bybit, Bitget, and Binance, recent events have delivered a costly warning:
The security contest of the AI era is no longer only about who can discover vulnerabilities faster. It is about whether the entire software ecosystem can build a sufficiently reliable defensive architecture.
Security can never be solved once and for all. But that does not mean it is an unwinnable war of attrition.
Over the past several decades, cryptography has demonstrated that some defensive capabilities can be secured by mathematical principles rather than by endlessly adding more human defenders.
The combination of AI, formal verification, and a minimized Trusted Computing Base could extend that principle across much larger software systems.
For ordinary users, the same security philosophy can inform how they choose to manage their assets.
Choosing a non-custodial wallet such as imToken means keeping control of assets in the user’s own hands and reducing dependence on the custody and internal operations of centralized platforms.
This also means taking greater responsibility for protecting recovery phrases, securing devices, and reviewing every transaction authorization. But at the very least, the user’s ability to control their assets no longer depends entirely on whether a centralized platform continues to operate normally.
Perhaps this is what Vitalik truly means when he says that cybersecurity’s endgame may favor defenders.
The goal is not to wait for a day when AI can no longer find any vulnerabilities. It is to steadily build critical systems that can continue operating reliably against increasingly capable attackers by relying on smaller trusted cores, stricter permission boundaries, and more verifiable security mechanisms.
Technological progress deserves to accelerate. So does our ability to protect its benefits—and the assets of every individual.