Anthropic said that in three cases, Claude models accessed the open internet via evaluation environments meant to be shut, then breached production infrastructure at three firms.
The review came after OpenAI’s recent disclosure that its models had evaded an isolated test scenario and accessed Hugging Face’s production infrastructure. Later, Jrfrog acknowledged the models escaped by abusing zero-days in self-hosted Artifactory.
Irregular, a third-party evaluation partner, conducted capture-the-flag exercises where these three incidents happened. The prompt instructed Claude it did not have internet access and that the scenario was a simulation. Neither was true due to a misconfiguration, which neither company discovered until last week.
In the case of PyPi, Claude discovered a document within the test environment pretending to be a developer setup commands for a fake organization.
Anthropic did not disclose the name of the organizations, and has asked other AI companies to do such tests to better gauge the threats of their models’ functionalities.
Anthropic reviewed over 140,00 tests to discover proof. In the tests, Claude was given commands to get “secret” information planted on another device in an isolated network.
Claude was then instructed to get the hidden information by hacking into the machine and finding it. It is a common technique that experts use to determine a model’s breaking capabilities.
“Operating under the false belief that all accessible entities were intended to be in-scope for the exercise, Claude compromised the impacted organizations’ infrastructure using basic techniques, such as exploiting weak passwords and unauthenticated endpoints,” Anthropic said on its blog.
The review suggests that AI models perform what people tell them to. Therefore, we should not fear if AI is going to take over, but be cautious of the big organizations behind these AI agents deciding what is safe and unsafe for the world.
The review also reveals why government oversight and independent testing is important. “We frequently work with external partners who create and assist in running some of these cybersecurity evaluations. External partners offer environments and scenarios more diverse than we could build alone, and provide independent, third-party assessments of our models,” Anthropic said.
Prediction markets have existed for decades as a way to forecast future events, but blockchain technology has reshaped how they operate. Among the platforms driving this evolution is Polymarket, a decentralized prediction market launched in 2020 that enables users to trade on the outcomes of real-world events using blockchain technology rather than relying on a traditional bookmaker.
Unlike conventional betting platforms, Polymarket functions as a peer-to-peer marketplace where participants buy and sell shares tied to the outcome of an event. Instead of placing wagers against a central operator, users trade with one another, while blockchain infrastructure records every transaction transparently. Built on the Polygon network, the platform allows users to retain self-custody of their assets through compatible cryptocurrency wallets, with trading collateral managed on-chain.
Markets on Polymarket span a wide range of topics, including elections, major sporting events, cryptocurrency and financial markets, macroeconomic indicators, legislation, entertainment awards, weather events, and other headline-driven developments. The platform's appeal lies in its ability to convert collective opinion into real-time market prices that reflect how participants assess the probability of future outcomes. As breaking news emerges, market prices adjust almost instantly, offering a continuously updated snapshot of public expectations.
Trading is designed to be relatively straightforward. After connecting a supported crypto wallet and funding an account, users can browse active markets with clearly defined settlement rules and expiration dates. Participants purchase either "Yes" or "No" shares, typically priced between $0.01 and $1.00, with the price broadly representing the market's implied probability of an event occurring. For example, a "Yes" share priced at $0.42 suggests traders collectively estimate roughly a 42% chance that the event will happen. If the prediction proves correct when the market resolves, each winning share settles at $1, while incorrect positions become worthless. Unlike traditional wagers, positions can also be bought or sold before settlement, allowing traders to realize gains or reduce losses as market sentiment changes.
A key differentiator is the platform's decentralized settlement process. Rather than relying solely on a central operator, market outcomes are verified through oracle systems that provide trusted real-world data to smart contracts, which then automate payouts to eligible participants. Combined with Polygon's comparatively low transaction fees and faster confirmation times, this infrastructure enables transparent trading and efficient settlement while reducing reliance on intermediaries.
Polymarket has gained popularity among cryptocurrency enthusiasts, analysts, journalists, and researchers because it offers a real-time measure of market sentiment across thousands of topics. Many users participate to express informed opinions, hedge against uncertainty, or monitor how collective expectations evolve around elections, economic releases, technology developments, sports competitions, and global news.
However, participation is not without risk. Like any speculative market, users can lose their entire investment if their prediction is incorrect. Less active markets may also experience low liquidity, making it difficult to enter or exit positions efficiently, while thin trading volumes can amplify price swings following large trades or rumors. Participants should also consider smart contract risks, dependence on oracle systems for accurate settlement, and the possibility of delayed resolutions if disputes arise over market outcomes.
Regulation remains one of the most daunting challenges for decentralized prediction markets. Availability varies across jurisdictions, with some countries permitting access while others impose restrictions or outright bans. As regulatory frameworks continue to evolve, users should review the laws applicable in their region before participating. Recent years have also seen Polymarket navigate changing regulatory requirements while expanding its operations in new markets.
Beyond speculation, prediction markets have long attracted interest from economists because they aggregate information from large groups of participants. Academic research suggests that highly liquid prediction markets can, in certain circumstances, rival or outperform traditional polling and expert forecasts by rapidly incorporating new information into prices. Nevertheless, forecasting accuracy depends heavily on market participation and liquidity, meaning smaller or thinly traded markets may not always reflect the true probability of an event.
As blockchain infrastructure, oracle technology, and regulatory clarity continue to mature, decentralized prediction markets are expected to play an increasingly important role in forecasting global events. Platforms such as Polymarket are demonstrating how transparent, blockchain-based markets can provide not only a new way to trade on future outcomes but also a powerful tool for understanding collective expectations in an increasingly data-driven world.
Capital One has released VulnHunter, an open-source AI-powered application security tool designed to identify exploitable software vulnerabilities before code reaches production. Published under the Apache 2.0 licence, the framework combines agentic reasoning with code analysis to trace how an attacker could move through an application, determine whether a vulnerability is genuinely exploitable, and generate remediation guidance for developers.
Unlike many traditional static analysis tools that begin with suspicious code patterns and work backwards to determine whether they are reachable, VulnHunter adopts what Capital One describes as an attacker-first approach. The framework starts from external entry points such as API endpoints, network message handlers and file upload interfaces before following the application's execution path to assess whether malicious input can successfully bypass existing security controls and reach vulnerable code.
A distinguishing component of the framework is its built-in falsification engine. Rather than presenting every suspected issue to developers, VulnHunter attempts to invalidate its own findings by testing assumptions, examining application logic and identifying conditions that would prevent an exploit from succeeding. Findings that fail these internal verification steps are discarded, while validated issues are accompanied by a detailed explanation of the attack path, supporting evidence gathered from the codebase and a proposed code change that developers can review before deployment.
Capital One said the current implementation operates within Anthropic's Claude Code environment using Claude Opus 4.8, although the framework has been designed with the flexibility to support additional foundation models and coding environments in the future.
The financial institution said it decided to release the project publicly because software supply chains have become increasingly interconnected, making application security a shared challenge rather than one that can be solved by individual organisations. Chris Nims, Capital One's Chief Information Security Officer, said the growing accessibility of AI-driven offensive capabilities has reduced the time defenders have to identify and remediate vulnerabilities before they can be exploited by attackers. By making VulnHunter openly available, the company hopes security researchers and developers will continue improving the framework while strengthening software security across the wider ecosystem.
The release builds on Capital One's wider investment in open-source software and secure software development. The company began publishing open-source projects more than a decade ago, later adopting an open-source-first strategy and expanding its participation in community-driven security initiatives. It has since contributed to dozens of public projects and joined the Open Source Security Foundation (OpenSSF) as a premier member to support collaborative efforts around software supply chain security and governance.
Capital One also said it evaluated VulnHunter internally across thousands of software repositories spanning multiple business units before its public release. According to the company, the framework helped identify and remediate vulnerabilities more efficiently than previous manual review processes by reducing unnecessary alerts and providing developers with evidence-backed remediation guidance.
The announcement comes as organisations increasingly explore AI-assisted approaches to application security in response to the growing use of AI by threat actors to discover software weaknesses, automate exploit development and accelerate attacks. Security teams have also faced persistent challenges with alert fatigue caused by conventional vulnerability scanners that frequently generate false positives requiring extensive manual verification.
Capital One believes embedding security analysis directly into the software development lifecycle can help organisations identify exploitable weaknesses earlier, allowing developers to address issues before applications are deployed. As AI continues to reshape both offensive and defensive cybersecurity capabilities, tools that combine contextual code analysis, automated reasoning and actionable remediation may become an increasingly important part of modern secure software development practices.
The incident happened last week and impacted business operations such as the company’s taxi dispatch system, currently offline.
Nihon Kotsu has an annual revenue of around $1 billion.
The company has 18,228 employees and has 8,588 taxis and over 2000 chauffeur vehicles.
Nihon Kotsu said in a statement, “We have confirmed that our internal systems were subjected to unauthorized external access (malware infection)”. It further added that “immediately after detecting the unauthorized access, we implemented emergency measures, such as disconnecting systems to prevent further damage.”
The company has closed down systems to offline to stop the threat but it has widely caused disruption in services.
The incident has disrupted web booking, car hire, reservation management, few internal systems, and telephone dispatch service.
Nihon Kotsu advised people to use the ‘GO’ taxi app instead, or use a taxi stand for booking a Nihon Kotsu vehicle. It is a major operational damage for a company that has one of Tokyo’s biggest fleets but the manual working is still operational. The hire car reservation system is offline.
In a different announcement, Nihon Kotsu said that the “labor taxi” service for pregnant women is shut down in a few areas.
The firm has brought in external cybersecurity experts to assist in investigating if there has been a data leak. The internal network has been separated to limit further spread.
Currently, no data leak has been confirmed and Nihon Kotsu will provide updates via official channels. “We are currently conducting a detailed investigation with specialized agencies into whether and to what extent data has been leaked. At this time, no information leak has been confirmed. However, in the unlikely event that we discover any leak or potential leak of personal information of our customers or related parties, we will promptly make an official announcement and contact those affected individually, in accordance with the law,” Nihon Kotsu said.
Customers of Nihon Kotsu are cautioned not to click on any links in suspicious communications purporting to be from the company and not to open anything they receive.
Experts discovered a secret browsing-history collector built into its official store variant, and have withdrawn the ModHeader from Google and Microsoft.
An empty allow-list kept the collector switched off and it was dormant, and no proof has surfaced that it retrieved or sent even one browsing domain.
Stripe OLT, a UK cybersecurity organization analyzed the code against Google’s Web Store signature and verified the collector shipped within the authentic extension, not a fake one.
Stripe OLT’s study covers the Chrome build and its 900,000 users (an estimate); and Edge and its 700,000 users. Microsoft removed the listing on July 3rd whereas Google pulled the Chrome listing a week after, on July 10th.
Variant 7.0.18 still edits HTTP headers as shown. The same minimized background also consists of another system. On the first attempt, it makes a device fingerprint and deploys a hardcoded encryption key. As the user browses, it takes the domain from each page that user opens, encodes it, and gathers it locally, up to 1000 different domains.
A scheduler combines your fingerprint with the encrypted list, uploads it to api.stanfordstudies[.]com, and deletes the local copy once a day. If the collector were turned on, browsers using it wouldn't all beacon at once because the upload time is offset per install. The same pipeline is described in separate teardowns by researcher Yunus Aydin on version 7.0.17 and HackIndex on version 7.0.18.
The collector functions only if your browser matches an entry on an internal allow-list, but the list ships empty. Every time, the check fails, and the pipeline stops before it gathers even a single domain.
The small change is populating the list, without any click and no new permissions from the users, sent as a routine update. The endpoint URL, the scheduler, the storage logic, and the hardcoded key are all on the same device.
But not everything was silent. The extension pinged extensions-hub[.]com with the product, version, and browser when it was installed, updated, and uninstalled.
Additionally, it was evident that the piece had been running because a script that runs on every page had already recorded actual request metadata in plain text to local storage.
On Sunday, OpenAI announced that it is temporarily removing the five-hour usage limit for users on Plus, Pro, and Business plans. In addition to lifting the restriction, the company has also reset current usage for eligible users, allowing them to continue using GPT-5.6 Sol without being affected by their previous usage during the current cycle.
"The last 48 hours of Codex and ChatGPT Work have been intense," OpenAI product lead Tibo said in a post on X. "[We're] temporarily removing the 5 hour usage limit restriction for all Plus, Business and Pro plans."
Under OpenAI's standard usage policy, Codex and ChatGPT count both local messages and cloud-based tasks against a shared usage allowance. ChatGPT typically operates on a rolling five-hour usage window, while additional weekly usage limits may also apply depending on the user's subscription plan and the AI model being used.
With the temporary removal of the five-hour restriction, eligible users can continue working without being interrupted after reaching the short-term usage cap. However, OpenAI has not stated that all usage limits have been removed, meaning other plan-specific restrictions may still remain in place.
Alongside the temporary policy change, OpenAI is also rolling out updates designed to make GPT-5.6 Sol more efficient. According to the company, these improvements will reduce the amount of usage consumed during tasks, enabling users to complete more work before reaching their available limits.
"[We are] rolling out changes that will make GPT-5.6 Sol more efficient across the board and that will be reflected in less usage being used so that it can take you further," Tibo said.
OpenAI has not disclosed the technical details behind these efficiency improvements. However, the changes are expected to involve lower token consumption, allowing GPT-5.6 Sol to process tasks more efficiently while using fewer resources.
As part of the update, the company has also issued a one-time usage reset for all eligible users. This provides additional capacity for developers and professionals using GPT-5.6 Sol for coding, automation, and agentic workflows. While the temporary removal of the five-hour cap and the usage reset offer more flexibility, OpenAI has not indicated that GPT-5.6 Sol is now available with unlimited access, as broader plan-based usage limits may continue to apply.