From Binance to PHP2Shell: Web Security Defenses Urgently Need to Be Redefined in the AI Era

Article by imToken

Early on October 30, the crypto market witnessed yet another historic coin theft incident, this time targeting Binance.

According to preliminary investigations, the attacker may have breached a DevOps engineer's laptop, accessed a service container, and submitted unauthorized withdrawal transactions. Binance officially estimated the total impact at up to $6.3 billion, and has currently co-frozen approximately $3 billion in related assets.

It goes without saying that over the past year, security incidents at top-tier exchanges have become increasingly frequent, leaving many feeling helpless:

Last year, Bybit suffered a theft of approximately $1.46 billion in assets; in September this year, Bitget had about $387.5 million in assets transferred without authorization; and now even Binance has been swept up in a new crisis.

Observing these recent security incidents together reveals a clear trend: attackers are less frequently trying to directly crack cryptography and private keys themselves, and instead are focusing on the entire "chain of trust"—from employee endpoints and third-party security products to software supply chains, service containers, and transaction authorization processes. A single crack in any link could cause the entire defense line to collapse.

The rapid development of AI is further amplifying these risks.

On October 22, Ethereum co-founder Vitalik Buterin offered a new assessment: In the AI era, the endgame of cybersecurity may favor defenders, but the transitional period leading to this outcome will be extremely difficult.

I. From Bybit to PHP2Shell: Attackers Are Piercing the Entire Chain of Trust

Let's first review these incidents.

Their attack vectors differ, but they share a common thread: attackers did not simply rely on cracking private keys or breaching underlying blockchain protocols. Instead, they exploited a component within the system that was inherently trusted by default.

In the Bybit incident, attackers compromised a Safe developer's device through social engineering, obtained relevant access credentials, and subsequently took control of frontend resources within Safe's wallet infrastructure;

In the Bitget incident, attackers exploited vulnerabilities in a third-party security product to obtain high-privilege internal credentials, then bypassed existing risk control mechanisms via forged withdrawal instructions;

In the latest Binance incident, the attack vector likely originated from a compromised laptop belonging to a DevOps engineer;

These three cases respectively involve third-party wallet infrastructure, third-party security products, and internal development and operations environments.

This also explains why traditional cold wallets and multi-signature mechanisms cannot solve all problems, since what attackers truly exploit are precisely those trust links capable of influencing final transaction execution yet lacking sufficient independent verification.

Similar chain-of-trust crises do not exist solely within crypto exchanges.

Just a week before the Binance breach, Binance founder CZ specifically tweeted to warn everyone to exercise extreme caution when visiting any website in the coming days—as two epic, high-severity vulnerabilities in PHP's core engine, PHP2Shell (CVE-2026-109618 and CVE-2026-109619, with a CVSS v4.0 score of 9.2), were sending waves of panic across the internet.

As an open-source server-side language powering approximately 70% of websites worldwide, mainstream platforms such as WordPress, Magento, and Drupal are all built upon PHP.

These two vulnerabilities, which had lain dormant for over a decade regarding memory management and global state reset, allow attackers to achieve Remote Code Execution (RCE) without authentication. With just a single specific web request, they can execute code before the site's own application logic runs, directly taking over servers, planting backdoors, or stealing user credentials.

Scanning data from cybersecurity firms Palo Alto Networks and Censys indicate that between 22.6 million and 36.7 million publicly exposed instances globally fall within the risk zone.

Even more staggering, according to disclosures from security agencies, these two deeply hidden memory state vulnerabilities, which had slept for over a decade and were extremely difficult for human auditors to detect, were actually uncovered using AI technology. In other words, AI has significantly lowered the threshold for converting concealed low-level logical flaws into usable attack payloads.

We have previously discussed multiple times the three qualitative shifts in cyber attack methods in the AI era (Further reading: When Hackers Use AI "More Efficiently," How Does the "Sword and Shield" Arms Race in Web3 Upgrade?):

The "industrialization" of zero-day (0-day) vulnerability mining: In the past, hunting for deep memory vulnerabilities in PHP core engines or firmware required top security experts to spend months on manual code audits and fuzz testing. Now, AI can scan global open-source code repositories around the clock at minimal cost, batch-converting decade-old latent bugs into lethal weapons;

The extreme compression of the "patch window" (N-day): Once vulnerability information leaks or patch code is submitted, malicious actors leveraging AI can reverse-engineer the patch, automatically generate variant exploits, and complete network-wide automated scanning and injection within minutes. Traditional IT operations teams' upgrade schedules, calculated in days or weeks, become utterly futile against millisecond-responsive AI attack scripts;

Software supply chains and DevOps becoming the primary breakthrough point: Modern Web3 and financial systems are not built in a vacuum. Frontend interactions, backend containers, third-party NPM/PyPI dependency packages, and engineers' local development environments form an extremely vast and fragile open-source supply chain;

Just as Binance's DevOps container was compromised this time and Bybit faced frontend supply chain tampering, given that on-chain attacks are the easiest to monetize directly, the crypto industry is experiencing what amounts to a "pandemic-scale wave of global AI hacker attacks."

Of course, not all the aforementioned attacks were directly launched or executed by AI. However, AI is simultaneously amplifying the capabilities of both attackers and defenders, and under the current software engineering framework, this amplification effect only exacerbates the structural asymmetry that already existed between offensive and defensive sides.

II. Vitalik: The Endgame May Favor Defenders, But the "Transition Period" Will Be Immensely Difficult

Amid industry discussions triggered by AI uncovering low-level vulnerabilities like PHP2Shell, Vitalik Buterin offered a highly measured assessment on October 22: The long-term endgame of cybersecurity may tilt in favor of defenders, merely that the transitional phase will be exceptionally challenging.

Why?

Because under the current software engineering paradigm, there is a natural 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 structures and execution paths.

Defenders must ensure every critical component is sufficiently secure, while attackers only need to find one undiscovered flaw.

Vitalik specifically noted that relying solely on security researchers to discover vulnerabilities in released software before hackers do does not truly establish a defensive advantage, because if both sides utilize increasingly powerful AI, it essentially boils down to a race to who can find vulnerabilities faster.

What truly needs to change is how software itself is constructed, exemplified by the two core directions he proposed: Formal Verification and minimizing the Trusted Computing Base (TCB).

The first is formal verification, aimed at replacing "experience-based debugging" with mathematical proofs.

Traditional software testing can only prove that "no bugs were found in tested scenarios," but can never prove that "the code absolutely has no bugs." Formal verification, however, uses rigorous mathematical logic to demonstrate that a piece of program code strictly adheres to expected security standards under all possible inputs and state transitions, eliminating flaws such as buffer overflows, unreinitialized states (like in PHP2Shell), or logical reentrancy.

In the past, the cost of formal verification was extraordinarily high, as writing mathematical proof code often required multiples of the business logic code volume, making it accessible only to aerospace or core cryptographic protocols.

However, in the AI era, AI can not only hunt for vulnerabilities but also act as a super-assistant for generating formal verification proofs. Thus, Vitalik predicts that "AI-assisted formal verification" will become the ultimate form of software development—whether for encrypted communication protocols like Signal, TLS, ZK-EVM, or application-layer smart contracts, end-to-end mathematical security proofs can be automatically generated by AI before code deployment.

When a piece of code is mathematically proven to be "flawless," no matter how powerful an attacker's AI computing power is, it cannot find vulnerabilities that do not exist.

The second approach is streamlining the Trusted Computing Base (TCB), aimed at compressing the attack surface to its absolute limit.

What is the Trusted Computing Base (TCB)? Simply put, it is the sum total of all hardware and software components that a system "must unconditionally trust to function correctly" to ensure security.

Reviewing Binance's $6.3 billion security incident and Bybit's hijacked multi-sig event: Why were assets lost when the underlying private key algorithm wasn't cracked? Because their TCBs were far too large—massive container clusters, complex operational privilege pipelines, developers' laptop operating systems, and frontend rendering JavaScript libraries were all incorporated into the chain of trust where "once breached, it leads to fund drainage."

The core philosophy of TCB minimization is subtraction and isolation: converging functions that genuinely involve private key protection, signature verification, and core authorization into smaller, clearly bounded, strictly auditable secure environments, while confining other complex functionalities to run in peripheral environments with restricted permissions whenever possible.

Even if browsers, plugins, frontend services, or AI Agents are compromised, they should not thereby gain the ability to unrestrictedly mobilize user assets.

From this perspective, this is also the value proposition of non-custodial wallets like imToken. Using centralized exchanges requires user assets to pass through platform accounts, custodial wallets, and internal withdrawal systems; whereas using imToken allows users to self-custody their private keys, directly manage on-chain assets, and initiate transactions, reducing reliance on the centralized custody intermediary step.

A shorter chain of trust means fewer platform-level risk exposures, eliminating concerns about exchanges halting withdrawals or freezing platform account balances. Of course, users must also bear the responsibility of safeguarding private keys and ensuring device security themselves.

III. From Technological Acceleration to Defensive Accelerism: Why We Need d/acc?

If formal verification and TCB minimization answer "how to become technically safer," then d/acc, long advocated by Vitalik, attempts to address a more macroscopic question:

When AI and other cutting-edge technologies are rapidly reshaping the world, how should we ultimately prioritize technological development?

d/acc stands for Defensive Accelerationism, commonly referred to as "defensive accelerationism." Vitalik first systematically introduced this concept in 2023 and further elaborated on it in his 2025 publication *d/acc: one year later*.

It does not oppose technological progress, but rather emphasizes a selective acceleration: prioritizing the development of technologies that enhance societal defensive capabilities, decentralize power, and reduce systemic risks.

Compared to e/acc's emphasis on blanket acceleration, d/acc focuses more on whether technology is actually bolstering offensive capabilities or strengthening defensive ones (Further reading: When Web3 Collides with d/acc: What Can Crypto Do in the Era of Technological Acceleration?).

Under the d/acc framework, this defensive architecture encompasses deep-building across four dimensions:

Cyber Defense: Reducing the success rate of cyberattacks and protecting digital assets and data through cryptography, formal verification, software isolation, and vulnerability detection;

Bio Defense: Enhancing society's capacity to respond to biological security risks through pathogen detection, air monitoring, vaccine research, and public health technologies;

Physical Defense: Mitigating the vulnerability of real-world critical infrastructure during attacks via distributed energy grids, resilient communication networks, and infrastructure hardening;

Information Defense: Countering misinformation, deepfakes, and mass information manipulation through privacy protection, source verification, and decentralized collaboration mechanisms;

Although these domains appear vastly different, they collectively revolve around a single objective: ensuring that as technology continuously empowers individuals and organizations, it also equips them with matching self-protection capabilities, preventing more powerful technologies from being monopolized by a few entities wielding resources and control.

It was precisely during the discussion surrounding this PHP vulnerability storm that Vitalik reiterated a hardcore stance under the d/acc framework—facing the asymmetric shocks brought by frontier AI, the unchecked advancement of high-risk offensive AI should be approached cautiously at necessary stages; meanwhile, all sectors of society should systematically allocate over 1% of GDP resources toward strengthening cybersecurity, biosecurity, and physical world security.

Why is such large-scale systematic investment necessary?

Because the benefits derived from technological progress can often be quickly priced and commercialized by the market, whereas the value of security infrastructure is frequently undervalued.

Companies are willing to invest massive capital into more powerful AI models, faster computing power, and more efficient software development tools, yet may hesitate to provide equivalent support for the long-term maintenance, formal verification, and security auditing of foundational open-source software. If this resource mismatch is not actively corrected, the digital edifice built upon fragile open-source foundations will only grow increasingly perilous.

Final Thoughts

From the global internet infrastructure risks exposed by PHP2Shell to the major security breaches successively impacting Bybit, Bitget, and Binance, the 2026 cybersecurity landscape is reminding everyone—at a literal cash cost—that:

The offensive-defense race in the AI era is no longer just about who can discover vulnerabilities faster, but fundamentally hinges on whether the entire software ecosystem can establish a sufficiently reliable defensive architecture.

Security can never be solved once and for all, but that does not mean it is a losing attrition war.

Over the past few decades, cryptography has demonstrated that certain defensive capabilities can establish robust security guarantees through mathematical principles, rather than relying solely on defenders pouring in more manpower.

The integration of AI with technologies like formal verification and minimized TCB also offers the opportunity to extend this philosophy further across entire software systems.

For ordinary users, this security concept can equally translate into choices regarding asset management methods.

Opting for a non-custodial wallet like imToken essentially keeps asset control firmly in your own hands, reducing reliance on centralized platform custody and internal operational frameworks. Naturally, this also implies that users must place greater emphasis on mnemonic phrase storage, device security, and every transaction authorization. At the very least, whether you can dictate your own assets no longer hinges entirely on whether a specific centralized platform is functioning normally.

Perhaps this is the true meaning behind Vitalik's assertion that the cybersecurity endgame may favor defenders: It is not about hoping that someday AI will simply fail to find any vulnerabilities, but rather about making incremental daily progress in building our mission-critical systems so that, even when facing increasingly sophisticated attackers, they can continue to operate reliably by relying on a smaller trusted core, stricter permission boundaries, and more verifiable security mechanisms.

Technological progress deserves acceleration, but the capability to protect technological achievements and individual assets needs to accelerate just as urgently.