Qubic says 87.1% of the drained QU remains immobilized on its network, while MEXC froze the account receiving the rest…. Read the full story at The Defiant
-
Claude Mythos and DeFi: Real threat or overblown fear?
When Anthropic introduced Claude Mythos-class models as its most advanced AI system for cybersecurity, it drew the usual mix of reactions from crypto communities. The lineup included Claude Fable 5, a Mythos-class model intended for broad use, although access was later suspended after a US government directive.
The concern around decentralized finance (DeFi) was easy to understand. If AI systems can find software flaws faster and with less human input, attackers may also use them to spot weak points in protocols before security teams can fix them.
Those concerns may seem overstated, but they come from a real shift in technology. AI tools have become better at reviewing code, spotting flaws and supporting security teams. At the same time, DeFi remains a major target for attackers because its code is often public, its protocols hold large amounts of money and many systems are new or not fully battle-tested.
The key question is whether Claude Mythos and similar tools pose a serious threat to DeFi, or whether the industry is overstating what today’s AI can actually do.
The answer sits somewhere between the hype and the alarm.
-
What is Claude Mythos?
Claude Mythos is Anthropic’s most advanced AI system for cybersecurity. Unlike general-purpose AI assistants that can write code or explain technical concepts, Mythos is designed to handle complex security tasks.
Anthropic initially limited access to the model instead of releasing it widely. According to the company, Mythos showed clear improvements in vulnerability research, exploit analysis and layered cybersecurity reasoning compared with earlier versions.
That capability drew attention quickly because vulnerability detection is valuable in both cybersecurity and crypto.
A security expert might spend weeks reviewing code for small flaws. If AI can shorten that timeline to hours, or even less, it could change the balance in defensive security.
That possibility explains much of the unease in crypto circles.
-
Why Claude Mythos matters to DeFi
DeFi has lost billions of dollars to hacks, exploits and protocol failures in recent years. The concern is not new.
Flash-loan attacks, cross-chain bridge exploits, governance attacks and smart contract bugs have shown that even audited protocols can still have gaps.
Unlike traditional software systems, DeFi protocols often control large amounts of money through smart contracts. A vulnerability may not just expose information. It could allow attackers to move funds quickly and without permission.
That makes DeFi especially attractive to malicious actors.
The open-source nature of many blockchain projects adds another risk. Their code is available for security teams to review, but it is also available to attackers.
In the past, finding advanced vulnerabilities required deep technical skill. Security researchers needed strong knowledge of coding languages, blockchain architecture, cryptography and attack methods.
AI changes that.
Instead of manually reviewing large codebases, analysts can now use AI assistants to flag suspicious patterns, summarize complex systems and point out possible attack paths.
This is where concerns around Claude Mythos begin.
Did you know? In some controlled security competitions, AI systems have identified software vulnerabilities in minutes that would normally take human researchers several hours, or even days, to find.
-
Can AI really find vulnerabilities in DeFi protocols?
The short answer is yes. AI systems have already shown that they can find certain types of software vulnerabilities.
Studies from Anthropic and other research groups show that advanced models can review code repositories, test security assumptions and sometimes find issues that human analysts miss.
Smart contracts are well suited to this kind of analysis because they are often public and written in structured languages such as Solidity.
An AI system can quickly review thousands of contracts, spot repeated patterns and look for known types of vulnerabilities.
Areas where AI is likely to provide growing support include:
- Reviewing audit reports
- Identifying unsafe coding practices
- Comparing protocol upgrades
- Detecting permission errors
- Modeling possible exploit paths
- Analyzing interactions between smart contracts
AI is becoming a force multiplier for security researchers. A task that once required a full team of experts could increasingly be handled by a smaller group of security professionals using advanced AI tools.
That is a meaningful change, not just marketing hype.
The table below shows how Claude Mythos compares with other models:
-
Why AI threats to DeFi may be exaggerated
Even with these advances, there is a clear difference between finding a vulnerability and stealing funds. Many crypto attacks involve much more than spotting a flaw.
Attackers often need to:
- Understand complex protocol mechanics
- Bring in significant capital
- Coordinate multiple transactions
- Exploit market conditions
- Manipulate liquidity
- Navigate governance systems
- Avoid detection
Even when a vulnerability exists, turning it into a successful attack often requires detailed planning and careful execution.
The real-world environment is far more complex than isolated coding tests.
Current AI systems also have limits. They can reach wrong conclusions, miss key details or follow weak lines of analysis. Security experts often find that AI tools produce useful insights alongside many false alarms.
An AI tool might flag 10 possible vulnerabilities, but only one may turn out to be valid. That matters because skilled human oversight is still essential.
Claude Mythos could speed up vulnerability detection, but it does not remove the need for experienced security experts.
Did you know? Many DeFi protocols publish their code online. This gives both security teams and AI tools more real-world financial software to review than in traditional banking systems.
-
The defensive side of AI in DeFi
A major flaw in the claim that AI will weaken DeFi is the idea that only attackers will benefit from these tools. Security teams have access to them too.
Security firms are already adding AI to their review processes. Developers are using AI-assisted code checks more often. Bug hunters can also use AI to spot issues before attackers find them.
Over time, AI may become a normal part of protocol security.
That could mean:
- Every code update goes through AI-assisted review
- AI agents continuously monitor deployed contracts
- Automated systems look for unusual on-chain activity
- Possible vulnerabilities are flagged before deployment
In that case, AI could strengthen DeFi security instead of weakening it.
The technology is neutral on its own. Its impact depends on how well attackers and defenders use it.
-
When AI attacks meet AI defenses
A more realistic outlook points to a future where AI systems challenge each other directly. This would make security faster on both sides.
Attackers will use more advanced models to find vulnerabilities and plan attacks. Security teams will use similar tools to monitor threats, improve code quality and respond faster.
This already happens in traditional cybersecurity, where offensive and defensive tools improve side by side.
DeFi could become the next major battleground for this contest. The likely result is not a sudden collapse of the sector. Instead, DeFi may enter a period of faster security upgrades and adaptation.
Projects that are slow to find vulnerabilities and update their code could face greater risk. Those that adopt AI-supported safeguards may become stronger than before.
Did you know? Several major crypto losses have come from compromised private keys, social engineering attacks or governance manipulation rather than flaws in smart contract code itself.
-
Assessing protocol vulnerabilities
Risk is not spread evenly across DeFi. Smaller projects with limited security resources often face the highest exposure.
Several categories are especially vulnerable:
- Fast deployment schedules: Projects that prioritize quick launches over careful testing may leave structural flaws in place.
- Copied codebases: Many protocols reuse or slightly modify existing code. Advanced AI tools can compare these systems quickly and expose inherited flaws.
- Weak audit coverage: Projects with little or no third-party review are less prepared for advanced attacks.
- Legacy smart contracts: Older contract designs may rely on assumptions that no longer hold up against modern exploit methods.
Automated analysis tools could sharply reduce the time needed to find these weaknesses.
-
What DeFi builders should do now
Claude Mythos offers an important lesson for the industry. DeFi builders should assume that attackers may already be using automated research tools. Security strategies need to improve accordingly.
Core priorities should include:
- Expanding automated security testing
- Running continuous, real-time audits
- Adding AI-assisted code analysis to development pipelines
- Increasing bug bounty rewards
- Using formal verification for critical code
- Improving threat monitoring and real-time incident response
Engineering teams must reduce the time between finding a vulnerability and deploying a fix. In an AI-accelerated environment, response time becomes just as important as prevention.
-
A major shift, not DeFi’s breaking point
Claude Mythos has shown that automated systems can handle complex security tasks that once required specialized experts. That marks a major shift for DeFi, where a code flaw can lead to the immediate loss of user funds.
Still, predictions of total systemic failure ignore several practical realities. Finding a vulnerability does not guarantee a successful exploit. Current AI tools still produce uneven results, human oversight remains essential and defensive teams have access to the same technology.
The more likely outcome is a change in security standards, not a collapse of DeFi. Automated tools could reduce the time and cost needed to find vulnerabilities. That will put more pressure on development teams to improve code quality, respond faster and build stronger security systems.
Ultimately, these developments are a warning, not a guaranteed outcome. The future of decentralized infrastructure will not be decided only by what AI can find. It will also depend on whether attackers or defenders use the technology more effectively.
The Raydium AMM V3 exploit drained roughly $1.34 million from a phased-out program tied to five pools outside the current product path, unsupported by Raydium’s UI or SDK, and inaccessible to current users.
The exploit hit legacy DeFi contracts and infrastructure that nobody treated as a live attack surface, exposing a lifecycle-management failure that extends well beyond one Solana decentralized exchange.
The category nobody is counting
Public exploit reports have found at least eight clear cases since March 2025 in which deprecated, obsolete, or legacy DeFi contracts became the attack surface, totaling roughly $10.8 million in losses.
Extending the definition to include broader legacy-vault and legacy-product failures lifts the count to about ten incidents and $22.5 million, including Raydium.
Exploit trackers classify incidents by technical mechanisms, such as smart contract bugs, access control failures, oracle manipulations, private key compromises, and bridge flaws.
Zombie contracts, or legacy DeFi contracts still callable after retirement, belong to a different axis entirely: a lifecycle state that consistently vanishes inside broader exploit labels.
| Exploit label databases usually use | What it captures | What it misses |
|---|---|---|
| Smart contract bug | The code flaw that let funds move | Whether the contract was deprecated, obsolete, or outside the active product |
| Access control failure | Missing or broken permission checks | Whether the affected deployment should still have been callable |
| Business logic flaw | Broken assumptions inside protocol logic | Whether the logic belonged to old infrastructure no longer supported by the UI/SDK |
| Oracle/accounting issue | Incorrect pricing, balances, or shares | Whether the vault or pool was a legacy product |
| Zombie-contract / lifecycle risk | Deprecated infrastructure still live on-chain | The missing category: contracts that were “retired” in product terms but not decommissioned technically |
Raydium’s AMM V3 pools were deprecated after Serum’s own deprecation rendered them inert. The legacy program was built to place orders on the Serum order book, and once Serum wound down, it lost its only function and left associated liquidity idle.
Raydium’s current programs use a virtual supply mechanism for proportion checks and verify LP mint addresses along with all other relevant account information.
The legacy program skipped both checks, letting an attacker create a new mint, present it as the LP token, and bypass proportion controls entirely.
Roughly 150,177 RAY, 5,603 SOL, and 893,700 USDC had been sitting in pools outside the current product but stayed callable on-chain.
One pattern for eight incidents
In March 2025, 1inch lost roughly $5 million when an obsolete Fusion v1 resolver contract implementation was exploited.
In October 2025, Abracadabra lost $1.8 million due to deprecated Cauldron V4 contracts that remained active and exploitable because of a logic flaw. In December 2025, Yearn’s legacy iEarn TUSD vault was drained of roughly $300,000, while Yearn’s current v2 and v3 vaults remained clean.
Things escalated in May: SlowMist reported Transit Finance losing $1.88 million through a deprecated 2022-era TRON contract, and Huma Finance lost roughly $101,000 through deprecated V1 BaseCreditPool contracts on Polygon.
Renegade lost approximately $209,000 due to a legacy V1 Arbitrum deployment exposed by an unprotected initializer and a migration issue, with white-hat recovery reducing the net impact.
Scallop lost roughly $140,000 due to a deprecated rewards contract, leaving the core lending infrastructure clean.
Every protocol made the same claim that current users were safe and current programs intact, and every protocol still paid out from the treasury, because the old infrastructure had stayed callable long after it left the active product path.
| Protocol | Date | Legacy surface exploited | Approx. loss | Why it fits the pattern |
|---|---|---|---|---|
| 1inch | Mar. 2025 | Obsolete Fusion v1 resolver implementation | ~$5.0M | Old resolver logic remained relevant enough to exploit after the protocol had moved on. |
| Abracadabra | Oct. 2025 | Deprecated Cauldron V4 contracts | ~$1.8M | Deprecated contracts remained active and exploitable through a logic flaw. |
| Yearn | Dec. 2025 | Legacy iEarn TUSD vault | ~$0.3M | Legacy vault was drained while current Yearn vaults remained unaffected. |
| Transit Finance | May 2026 | Deprecated 2022-era TRON contract | ~$1.88M | Old contract surface stayed live after deprecation and became the attack path. |
| Huma Finance | May 2026 | Deprecated V1 BaseCreditPool contracts on Polygon | ~$0.101M | Retired architecture still held exploitable value outside the current system. |
| Renegade | May 2026 | Legacy V1 Arbitrum deployment | ~$0.209M | Migration and initializer issues exposed an old deployment. |
| Scallop | 2026 | Deprecated rewards-side contract | ~$0.14M | Core lending infrastructure stayed clean, but old rewards infrastructure was exploitable. |
| Raydium | 2026 | Legacy AMM V3 pools | ~$1.34M | Current UI/SDK and users were unaffected, but old pools remained callable on-chain. |
Why databases lose this
Most exploit classifications focus on how the attacker got in, what they manipulated, and which code failed, a mechanism-first lens that obscures zombie contract exploits, where the core failure is that the infrastructure was supposed to be retired.
Transit’s deprecated TRON contract was an old protocol surface that nobody decommissioned. Scallop’s deprecated rewards contract was an accounting flaw in infrastructure that the team had moved past. Huma’s V1 BaseCreditPool was retired architecture still holding assets on a chain the protocol had migrated away from.
A 2025 SoK paper analyzing 50 severe real-world exploits from 2022 to 2025, totaling over $1 billion in losses, argued that high-impact incidents frequently involve exploit chains spanning human, operational, economic, lifecycle, and governance layers.
The authors proposed a four-tier root-cause framework that treats lifecycle and governance failures as a distinct category alongside implementation errors. Zombie contracts fit that framework: lifecycle failures that exploit databases are absorbed into implementation-bug counts, keeping the cumulative dollar figure buried inside unrelated categories.
The fork in the graveyard
If protocols continue to treat decommissioning as an afterthought, deprecating contracts in product documentation without draining, pausing, or monitoring them, attackers will keep scanning the graveyard.
Every major protocol’s deployment history becomes a searchable attack surface. The $22.5 million current estimate is a floor, based on incidents that made it into public reporting with sufficient detail to classify.
Legacy vaults, forgotten approval surfaces, and old integrations that still hold assets but sit outside active user flows receive far less monitoring than live infrastructure, which is what attackers scan for.
If the category gets named and counted, if decommissioning checklists become standard practice alongside audits, the attack surface shrinks through maintenance.
Raydium’s treasury absorbs the $1.3 million exploit, Transit’s team promised compensation, and Huma covered its losses.
That makes DeFi contract decommissioning a security control rather than a documentation task.
| Decommissioning control | What it means | Why it matters |
|---|---|---|
| Drain idle assets | Remove funds from retired pools, vaults, and reward contracts. | Eliminates the financial incentive for attackers to scan abandoned infrastructure. |
| Pause callable functions | Disable swaps, withdrawals, reward claims, or admin functions where possible. | Turns “deprecated” into an actual security state rather than a product label. |
| Verify LP mints, approvals, and permissions | Review old mint checks, approvals, authorities, and account assumptions. | Prevents attackers from exploiting stale validation logic or forgotten permissions. |
| Monitor legacy deployments | Keep alerts active for old contracts, pools, and chain deployments. | Prevents abandoned infrastructure from becoming invisible to the team but visible to attackers. |
| Keep legacy code in bug-bounty scope | Include retired or deprecated infrastructure in security programs. | Gives white hats a reason to report issues before attackers exploit them. |
| Publish retirement status | Clearly identify whether old products are drained, paused, monitored, or unsupported. | Helps users, integrators, and analysts distinguish “not in the UI” from “not risky.” |
| Define treasury liability | State whether the protocol will compensate losses from retired infrastructure. | Makes clear whether old code remains an implicit claim on the protocol treasury. |
Deprecating a contract transfers the security liability to the treasury while leaving the attack surface intact. Retiring infrastructure without decommissioning it keeps it live, with the team’s attention diverted and the attacker’s incentive intact.
In addition to total value locked, DeFi protocols accumulate history, and history can be exploited.
