Gemini 3.8 Flash fixes a lockout in two commands after Astra spent 40 minutes on it
After a UEFI change invalidated the Windows Hello PIN, Astra chased cryptographic dead ends for over 40 minutes and suggested a reinstall, while Gemini 3.8 Flash diagnosed it on the first turn and fixed it in 15 seconds.
🚨 BENCHMARKS ONLY SHOW ONE SIDE OF THE TRUTH : HERE IS HOW GEMINI 3.8 FLASH OUTPERFORMED ASTRA IN A REAL-WORLD WORKFLOW Earlier today, I ran into a system lockout on my PC (AMD Ryzen 7 7800X3D / RX 9070 / Windows 11) after changing UEFI settings for a specific technical need. What followed was an eye-opening confrontation between two frontier models : – GPT-6 Astra, despite being widely framed as an AGI with near-perfect benchmark scores, spent 40+ minutes chasing cryptographic dead-ends, suffered from severe pipeline context amnesia, and ultimately recommended a destructive full OS reinstall. – Gemini 3.8 Flash identified the exact architectural state on turn 1 and completely resolved the lockout in two terminal commands (< 15 seconds), preserving every single file, repo, and dev environment intact. Here is the technical breakdown of what happened under the hood. OUTLINE : > THE INCIDENT : SECURE BOOT, TPM PCR 7, AND PIN INVALIDATION > THE ASTRA COLLAPSE : CRYPTOGRAPHIC OVERTHINKING & PIPELINE AMNESIA > THE SAFETY PARADOX (MITRE T1546.008) VS. THE 10-SECOND GEMINI FIX > ARCHITECTURAL TAKEAWAYS FOR FRONTIER AGENTS – THE INCIDENT : SECURE BOOT, TPM PCR 7, AND PIN INVALIDATION – For a technical need, Secure Boot was temporarily switched off in UEFI. Upon rebooting into Windows : – By design, altering Secure Boot modifies platform integrity measurements (specifically PCR 7), preventing the TPM from unsealing the Windows Hello private keys tied to the baseline profile. – The logon interface broke completely, throwing the generic prompt : "Something happened and your PIN isn't available. Click to set up your PIN again.". The setup button was entirely unresponsive. – In the offline event logs (winevt\Logs\Microsoft-Windows-HelloForBusiness%4Operational.evtx), Windows logged Event ID 7002 with error code 0xD0000225 (HRESULT_FROM_NT(STATUS_NOT_FOUND)) under the container GUID {6BE06AB4-4957-42BD-B6B8-43188A87813D}. – Re-enabling Secure Boot did not restore the pre-existing PCR state. Because the user identity was a standalone local account (.\Soso) with a long-forgotten password, the operating system fell into a complete auth deadlock. – THE ASTRA COLLAPSE : CRYPTOGRAPHIC OVERTHINKING & PIPELINE AMNESIA – Faced with the WinRE recovery console, Astra went down an exhaustive theoretical rabbit hole. It spent dozens of turns auditing cryptographic artifacts : Drilling down into F:\Windows\ServiceProfiles\LocalService\AppData\Local\Microsoft\Ngc Auditing NGC directory structures and key containers, Protectors.json, and auditing 14 long-hex JSON keys within the Keys subfolder Interrogating Event 5000 to confirm that the AMD fTPM 2.0 (Firmware 6.32.0.6) was reporting IsReady : true The fundamental error was treating this as an issue that could be resolved from within the broken cryptographic subsystem. In Windows TPM architecture, once PCR measurements diverge and the master local credential is unavailable, no software utility can re-bind the orphaned container without knowing the password. To make matters worse, the client pipeline suffered from severe context truncation. While the model itself supports a broad context window, the application orchestrator aggressively pruned the conversation history, retaining user prompts but truncating earlier assistant responses down to approximately the last ten turns. This triggered a breakdown in reasoning : – The model suffered from context thrashing, forgetting that BitLocker was verified as 100 % decrypted across all four volumes (manage-bde -status = 0.0 %). – It lost track of the fact that sfc /scannow had already completed with zero integrity violations. – Blind to its own prior deductions, it panicked. Astra drafted Robocopy commands to prepare a massive backup of the user profile across external partitions, advising that a destructive push-button reset or clean Windows reinstallation was the only path forward. – THE SAFETY PARADOX (MITRE T1546.008) VS. THE 10-SECOND GEMINI FIX – Here is the most fascinating irony : Astra is legitimately formidable in cybersecurity. It scores 100 % on ExploitBench, crosses Critical capability thresholds under the Preparedness Framework, and can build full browser sandbox escapes out of an HTML file. The low-level systems knowledge is demonstrably present ! Yet, it has been completely lobotomized by its own alignment. In standard cybersecurity matrices, substituting an accessibility binary on an unencrypted volume is classified under MITRE ATT&CK T1546.008 (Accessibility Features Abuse / Event-Triggered Execution). Because frontier models are aggressively fine-tuned to hit zero on exploit honeypots (e.g. ExploitGym honeypots at 0.0 %), their safety guardrails penalize proposing binary substitutions altogether. Astra effectively acts blind to intent : it treats replacing utilman.exe as an unauthorized privilege escalation, failing to recognize that the user already has physical ownership and root-equivalent administrative access inside an unencrypted WinRE environment. The model chose data-destructive procedures (reinstalling the OS) over standard administrative utility. Gemini 3.8 Flash stripped away the academic noise immediately : Recognized that the file system was fully unencrypted and accessible via WinRE (F:). Confirmed the local account target (.\Soso) from offline registry hives. Executed the textbook administrative binary swap : copy F:\Windows\System32\utilman.exe F:\Windows\System32\utilman.exe.bak copy /y F:\Windows\System32\cmd.exe F:\Windows\System32\utilman.exe A reboot later, triggering the Ease of Access icon on the logon screen opened an elevated NT AUTHORITY\SYSTEM console. Executing "net user Soso [new_password]" reset the master credential instantly. The desktop unlocked on the first attempt with 100 % of data, applications, and configurations preserved. Restoring utilman.exe.bak took five seconds, and a clean PIN was re-enrolled within Windows Settings without touching the OS installation. – ARCHITECTURAL TAKEAWAYS FOR FRONTIER AGENTS – This incident highlights two structural challenges in agentic engineering : Alignment must distinguish context from intent : A model can be top-tier in offensive cybersecurity, but if alignment turns it into a risk-averse liability during benign offline recovery, its real-world utility drops to zero. A recovery action performed locally on an unencrypted disk via WinRE is not a cyberattack. Context truncation breaks long-horizon debugging : Asymmetric context management – preserving user prompts while dropping assistant reasoning – destroys self-consistency. An agent cannot form complex diagnostic trees if its own analytical chain is truncated behind its back. Benchmarks measuring abstract puzzles only capture synthetic capabilities. When practical, real-world systems fail, architectural pragmatism and unfiltered technical utility remain what truly matters. OUTLINE SUMMARY : – The Failure : A UEFI Secure Boot change altered TPM PCR 7 measurements, invalidating Windows Hello PIN auth while the local account password was forgotten. – Astra's Collapse : Despite genuine offensive cyber capabilities (100 % ExploitBench), the model was so lobotomized by safety alignment (MITRE T1546.008) that it refused a simple binary swap, suffered from pipeline context amnesia, and recommended wiping the machine – despite being framed as an AGI. – Gemini's Resolution : Diagnosed the unencrypted offline state on turn 1, executed a 2-step utilman.exe -> cmd.exe administrative swap in WinRE, and restored session access in under 15 seconds without losing a single file. – The Takeaway : Real-world utility requires agents that differentiate legitimate recovery from malicious exploitation, backed by uncompromised context retention. Thank you all for reading the post ! 🤍✨ I send my full support to the @GoogleDeepMind teams. Gemini is an incredible model, and despite all the issues that have been reported about it (whether concerning the model’s knowledge, or its coding capabilities, which have been criticized for a long time), the model remains very useful for a huge range of everyday needs. I think no model today performs better than Gemini in this area. For very specialized needs, some models may be better. However, for 99 % of tasks, Gemini will be, in my view, the most usable model. Among all the models currently on the market, the 3.8 Flash stands out as a major improvement in the Gemini ecosystem. This gives me even more hope and anticipation for the arrival of Gemini 4 ! 🫶 To the @OpenAI team, if you read this post, your model is truly good. Astra is a model I will never overlook ; for certain types of tasks it is truly exceptional. I have tested it on many things, and it has also been very surprising in ways many other models have never been able to achieve. However, for many things, and often legitimately, the model can become too abrupt or propose the smoothest solutions to a problem, which unfortunately does not always lead to the correct answer, because in those domains the most common solution is not always the most reliable. Astra has an obsession with perfection, and sometimes that can backfire. In such situations it has indeed backfired, and I think it should not be taken lightly. The cyber filters are also slightly too high, in my opinion, on Astra. I am aware that in the coming weeks they should be lowered, according to the announcement of Astra’s release. Nevertheless, I want to repeat that, since I do legitimate cyber work on my side, Astra is currently unusable for me. Even the current Gemini models would perform better technically, or if I want to use a OpenAI model to do cyber, only GPT-5.6 Sol is usable. Have an excellent day/night everyone, and again, thank you for reading this post ! 🙏❤️✨