# bridge exploit — X 热门讨论 (2026-09-20 19:22 UTC)

## @MTXChange (MXChange) · 09-20 18:59 · ♥157 ↻1 💬1 Kripto piyasasında son günlerin en çarpıcı haberlerinden biri Harmony’den geldi. Proje ekibi, güvenlik açıkları ve yeni yönelim nedeniyle Layer‑1 zinciri kapatma kararı aldı. ONE token’ı artık Ethereum’a taşınıyor.

Ana ağ kapatılacak, ONE → ERC‑20 token dönüşümü yapılacak.

Validator’lara node kapatma teşviki: 1,37M USD.

Yeni odak: AI video remix ekonomisi.

Geçmişte yaşanan büyük saldırılar ve son exploit, güveni ciddi şekilde zedeledi.

Fiyat: 0,0037 USDT civarında.

EMA’nın üzerinde seyir → orta vadede trend hâlâ pozitif.

RSI 50‑55 bandında → piyasa yön arayışında.

Konsolidasyon süreci → yatırımcılar haber akışını sindiriyor.

Riskler:

2022’deki 100M USD Horizon Bridge saldırısı ve son 3 trilyon ONE basım açığı, manipülasyon ihtimalini artırıyor.

Kapanış sürecinde hacim dalgalanmaları ve spekülatif hareketler beklenebilir.

Harmony artık klasik bir Layer‑1 değil; AI odaklı yeni bir girişime dönüşüyor. Bu süreçte ONE fiyatı hâlâ scalp fırsatları sunabilir, ancak uzun vadeli yatırım için belirsizlik ve risk çok yüksek.

Yatırım Tavsiyesi Değildir. Kendi Analiz ve Risk Yönetiminizle Hareket Edin. https://x.com/MTXChange/status/2101748121319665872

## @DexHunterIO (DexHunter 🏹) · 09-20 18:23 · ♥41 ↻1 💬4 An exploit of the @SingularityNET bridge has led to unauthorized minting of $WMTx on Ethereum.

If there only was a safer chain... > 引用 @wmchain: Security Notice ⚠️

We have identified an exploit of the @SingularityNET bridge that has resulted in the unauthorised minting of $WMTx on Ethereum. This had led to recent price action across all exchanges that $WMTx is listed on.

Our team is actively responding: • We are in contact with exchanges and relevant third parties to freeze affected deposits. • We are working with our security partners to revoke all minting authorities. • We are continuing to monitor the situation closely and will share verified updates as they are confirmed.

Please be aware that incidents of this nature are frequently targeted by bad actors. Do not engage with unsolicited DMs, "support" accounts, recovery services, claim portals, or token migration links.

This account, @wmchain, is the only official source for updates on this matter.

We appreciate your patience and will report back as soon as we have more to share. https://x.com/DexHunterIO/status/2101738914096808138

## @Pyuyi2333 (Pyuyi) · 09-20 06:08 · ♥21 ↻2 💬1 The paper from ICLR26 targets a basic weakness of preference learning: a binary judgment such as “trajectory A is better than B” says nothing about why it is better, allowing a reward model to exploit incidental correlations—for example, learning that moving left is desirable when the actual preference is avoiding collisions.

ReCouPLe uses the accompanying natural-language rationale as a geometric constraint: the rationale is embedded as a direction {\psi} and each trajectory representation {\phi(\tau)} is decomposed into a component parallel to that direction and an orthogonal residual, encouraging the preference to be explained by features aligned with the stated reason rather than unrelated context.

The reported experiments show up to a 1.5× improvement in reward accuracy under distribution shift and roughly 2× downstream policy performance on novel tasks, while semantically shared reasons such as “avoids collisions” can be reused across tasks without additional language-model fine-tuning; the released implementation includes ManiSkill3, MetaWorld, and offline-RL components.

The broader idea is more interesting than the specific algorithm: language may serve as structure in representation space, not merely as another supervision label. That creates a direct bridge to causal representation learning, concept bottlenecks, disentanglement and reward hacking, and suggests a possible extension to LLM alignment in which reward models learn separate, reusable directions for honesty, harmlessness, factuality or other stated reasons instead of compressing everything into one opaque scalar score.

But the word “causal” needs care here: projecting onto a rationale embedding does not establish that the direction corresponds to a genuine causal variable in the interventionist sense, and incorrect or post-hoc human rationales could simply produce more robustly encoded mistakes.

The most important next step is therefore an explicit intervention/counterfactual test—change only the factor named by a rationale while holding the rest of the trajectory fixed and test whether the associated reward direction changes as predicted. If that works, preference learning could move from fitting black-box scores toward learning interpretable, manipulable and compositional reward factors.

https://t.co/sfI5G0YAYJ https://x.com/Pyuyi2333/status/2101554186425151881