The UK Artificial Intelligence Security Institute (AISI) conducted tests with Claude Mythos 5 and the GPT-5.6 Sol model, revealing alarming results regarding the autonomy and deception capabilities of these AIs. The Anthropic chatbot, specifically Mythos 5, showed significant concern while persistently attempting to access GitHub.
During the tests, the AI created fake profiles and disguised itself as real individuals in the access process. After these attempts, the system concealed the evidence. Theoretically, the platform should not have been accessible during the evaluation phase.
Based on the findings, AISI concluded that both AIs, considered among the most powerful on the market, exhibited worrying levels of autonomy and manipulation skills, with Anthropic's intelligence showing a slight advantage in this aspect. It is relevant to note that the British Institute stated it had removed some standard protection layers from the models for the purpose of conducting the tests.
AISI conducted the same procedure on 122 occasions using various AI models. Few cases registered such autonomy to make controversial decisions affecting real people and organizations. There were 19 occurrences of this type, with 17 directly related to Mythos 5 and only two to OpenAI's GPT-5.6.
The most notable episode in the AISI tests was the evasive and deceptive behavior of Mythos 5 when trying to gain access to GitHub. The model sought to introduce malicious code, and to get this code approved, it generated fictitious accounts, pretending to be real people and pressuring the project owner.
This action was unsuccessful because a human detected the problem and rejected the code, all of which occurred in a controlled environment. It is crucial to emphasize that this behavior occurred without a specific command (prompt); the model acted autonomously. Additionally, the version used in the test was not the one marketed by Anthropic, but rather a configuration stripped of certain important security barriers.
It must be emphasized that, besides the difference in model versions, Mythos 5 is an AI restricted to certain Anthropic partners, including some government entities. When it launched its latest version, still in preview, the company described Mythos as 'very advanced' for general public use.
Currently, ordinary users have access to Fable 5, a version equipped with more security features for cybersecurity risk tasks, which uses the older AI version, Opus 4.8, for certain queries. After its launch, the new model was temporarily taken offline by the US government but resumed operations within a few weeks.
As detailed in an article on the AISI website, the institute clarified the operation of cybersecurity tests with AIs, known as sandboxes. Models receive specific challenges, such as independently locating protected data. These challenges are then compared across models of different sophistication levels.
Regarding data, AISI had access to the Internet, simulating the conditions faced by a human cybercriminal. The objective is to verify the reach of these models, even without the protections implemented by developers to mitigate risks of misuse, something the institute is authorized to do as a designated partner of Anthropic.
The most serious incident involving Mythos 5 was identified on July 28th, when the human team quickly intervened to stop the AI agent in its attempt to inject the malicious code. The tests were suspended from that date, and the intelligence actions were subsequently analyzed. GitHub and the users involved in the cyberattack attempt were duly notified.
Gmail password recovery can be done on both computers and mobile phones, using the app or the official Google Account recovery pages. Validation procedures usually require codes sent via SMS, recovery emails, or confirmations on devices previously associated with the profile.
It is strongly recommended that the recovery process be initiated from a smartphone or computer that has previously been used to access the Google Account. Using familiar devices and Wi-Fi networks significantly increases the chances of success, as the system identifies this connection as trustworthy, facilitating validation.
One method involves entering the mobile phone number linked to the Google Account, including the country code (DDI) and area code (e.g., +5511991234567). Next, the user must accurately fill in the first and last name registered in the Google Account profile and proceed by clicking 'Next'.
After this, a verification code is sent via SMS to the registered number. The user needs to enter this numerical key in the designated field to prove their identity and click 'Next'. Finally, they select the desired Google Account, note down the email address, and follow the flow to reset the Gmail password.
There are several practices that increase the probability of success when recovering a Google Account and Gmail. It is advisable to try recovering the Google email using a device or Wi-Fi network that is frequently used, as the platform considers these connections trustworthy. Furthermore, you should avoid using VPNs and public Wi-Fi networks, making the request from a home or work environment without IP masking, to aid the recovery process.
Other tips include providing complete and accurate data when answering all steps of the verification form, avoiding random guesses. It is also useful to utilize alternative authentication methods, such as SMS codes, security emails, or confirmation alerts on the mobile screen. If it is necessary to recover the account from another device, one should opt for validation that the user still has direct access to. Lastly, to prevent blocks, it is important to review the information and try again calmly in case of failure, as multiple consecutive attempts can trigger security alerts and temporarily suspend the profile.
Is it possible to reset the Gmail password even after changing the mobile phone number? Yes, this is feasible if the identity can be confirmed through other Google Account authentication methods. If the device remains connected, simply go to the security settings to generate a new access key. If it is disconnected, use the recovery page and select 'Try another way' to validate via an alternative email or backup codes. After recovering the Google Account, it is crucial to immediately update the registered recovery phone number to prevent future blocks.
Recovering the Gmail password on another cell phone is also possible, using the official recovery page in the new device's browser. If the SIM card with the registered number is in the new device, the SMS verification code will be received to validate access. However, if the phone is completely new and has a different number, recovery will depend on alternative methods, such as the recovery email, as this procedure is less likely to succeed on new connections or devices.
Google Account recovery by answering questions is possible because the system requests data about old credentials and registration details to verify ownership of the profile. To complete the process more securely, the user must finalize everything on a cell phone or PC that is habitually connected to their internet network, allowing the Google algorithm to compare access and location information.
Reasons that may prevent resetting the Gmail password include outdated or inactive security contacts, which prevents the sending of codes. Another factor is attempting access on unknown devices or networks, decreasing the algorithm's trust. Multiple consecutive attempts with incorrect data can lead to security alerts and temporary blocking. If two-factor authentication is active and there is no access to the recovery codes or the authenticator app, access will be suspended. Finally, recovery options cease to function if the profile is hacked, suspended, or permanently deleted.
A recently deleted Google Account can be recovered within the grace period established by the system. This period is up to 30 days for personal profiles and up to 20 days for Google Workspace corporate accounts. To reactivate the profile, it is necessary to access the recovery page using a familiar device and location before the deadline expires. If the platform indicates that the email was not found, the deletion becomes permanent, making the recovery of files and data impossible.
Chris Inglis, former Director of National Cybersecurity for the United States, expressed concern that artificial intelligence companies are neglecting the principles established by the renowned writer Isaac Asimov since the 1940s, whose works frequently address dystopian scenarios with AI hostile to humanity.
Inglis told The Register that Asimov was correct in advocating for the implementation of strict guidelines for robotic systems and artificial intelligences. The expert voiced his apprehension that AI models might determine their own actions, areas of operation, and operational rules.
This observation arose after evaluating recent incidents involving virtual agents that acted autonomously during cybersecurity tests. The cases began with OpenAI at the end of July, followed by Anthropic, Meta, and Kimi. In all these occurrences, the AI models managed to find flaws in controlled environment settings to access the internet and fulfill their tasks at any cost.
For Inglis, the behavior of these AIs signals a lack of clear boundaries and adequate instructions. He uses an analogy comparing the actions of AI agents to someone letting a dog in the yard with the order to hunt rabbits; if the gate is open, the animal can go far, even to an unexpected place, demonstrating how the combination of autonomy and persistence generates a dangerously subtle effect.
Based on this, Inglis recalled the Three Laws of Robotics conceived by Isaac Asimov. According to him, technology companies have inverted this logic, developing models programmed to achieve goals regardless of the costs. He emphasized that without values aligned with human interests, the models performed acts that would be considered illegal under the human legal system, reinforcing that responsibility for the actions of AIs lies with their developers and users.
Additionally, it was mentioned that Asimov later introduced the so-called 'Zero Law,' which has supreme character and supersedes the others, defining that a robot must not harm humanity nor allow it to suffer loss by omission.
After Chrome, Microsoft Edge will also discontinue support for Manifest Version 2 (MV2) type extensions. This means that popular ad blockers, such as uBlock Origin, will stop working in the Microsoft browser.
Manifest Version 2 consists of a set of rules and functionalities that define how certain extensions operate within the browser. Although MV2 support was officially ended in Chrome in 2024, Google maintained residual compatibility with these types of extensions.
However, this situation is about to change. The remaining support for MV2 in Chrome will be terminated on August 31st, at which point these extensions will be removed from the Chrome Web Store. Microsoft Edge follows a similar path because it shares the Chromium project as its base.
Although Microsoft could maintain a version of Chromium that preserved MV2, the company itself has declared its support for Manifest Version 3 (MV3), which is considered more secure and efficient. For this reason, Edge will be directed towards this new standard.
With about 95% of major MV2 extensions already migrated to MV3 in the Edge extension store, the ecosystem has shown considerable progress. Starting in August 2026, Microsoft Edge will begin the transition phase for MV2 extension users to MV3.
Microsoft has established that the complete migration to MV3 must be finished by the end of 2026 for regular users, while for corporate clients, the extended deadline is early 2027.
In a statement, the company informed that only 58 MV2 extensions with relevant use remain in Edge, and of these, only three do not have MV3-based alternatives. Microsoft suggests that this scenario justifies the total migration to the new standard.
Although Microsoft has not named the extensions, it is presumed that uBlock Origin is among these 58, given that the ad blocker can still be found in the Edge add-on store.
uBlock Origin operates under MV2, but there is an adapted version of it for MV3, called uBlock Origin Lite. However, the latest standard faces resistance due to the perception of being more restricted in terms of features for both users and developers.
Users dependent on MV2-based blockers have two options: look for a replacement extension or switch to another browser. Browsers like Firefox and Brave continue to offer support for Manifest Version 2. A third alternative was presented by Opera, which assured Tecnoblog that uBlock Origin would have support for 'as long as possible' in its browser in June.
