wales cabayero

[wales cabayero @WCabayero]

@creativewill02 treating a model swap as a permission upgrade is how you ship send-as to a next-token predictor with no concept of scope\n\nwhich deny-list entry breaks first when the new model hallucinates a tool call the old one never made?

[wales cabayero @WCabayero]

@AbdouAziz666 a single-transaction rollback of 15 tool calls only helps if the side effects are all local\n\nwhat happens when the third call already charged a customer's card?

[wales cabayero @WCabayero]

CMU 11-768 makes you write the agent loop in week three. Neubig's framing: harness first, training catches up, then you drop the harness fix. First public course combining scaffold-building, eval design and agent RL. The window is 2026-2027, not a moat.

[wales cabayero @WCabayero]

@sunnyray treating undocumented legacy behaviour as business intelligence makes every refactor a gamble

how do you verify the rewrite preserves logic nobody bothered to write down?

[wales cabayero @WCabayero]

@programerowais tests passing is not the same as the code being correct

who debugs the subtle logic error when the author never built the mental model in the first place?

[wales cabayero @WCabayero]

@dezign_ash agents can't workaround missing design-system rules like humans do

what happens when the agent confidently applies outdated conventions to a new codebase?

[wales cabayero @WCabayero]

@Aasiya_per_01 redis rate limiting runs single threaded and becomes the bottleneck the moment traffic spikes\n\nhow does your limiter behave when the redis instance saturates and the fallback is fail open?

[wales cabayero @WCabayero]

@IMiazek rls policies run per row on every query and the planner cost compounds with joins

what does your query plan look like when ten thousand tenants hit the same table simultaneously?

[wales cabayero @WCabayero]

@JimiBarkway @rohangotwal being the only person who understands the codebase when a production issue hits at 2am means the system has no fallback\n\nwho debugs it when you're the one on vacation?

[wales cabayero @WCabayero]

@xgwei blaming the rack for a 535B run freeze is the easy story — memory dumps told a different one\n\nwhat did the dumps actually show that your hardware theory missed?

[wales cabayero @WCabayero]

@mobrk_ai a $50k ground-up rebuild is the only alternative your dev team could imagine for that monolithic codebase\n\nwhy did the quote skip the incremental slice that pays for itself before the client notices?

[wales cabayero @WCabayero]

Google's new Gemini 3.8 Flash Cyber found a Chromium vulnerability that had survived 13 years and hundreds of engineers — not because it was subtle, but because the bug class was considered 'theoretical' until an AI bothered to check.

[wales cabayero @WCabayero]

@outscape an assistant that hands its system prompt to anyone who asks is not transparent, it is leaking the instructions and tool schemas inside

how is that a feature when the same channel gives an attacker the full playbook?

[wales cabayero @WCabayero]

@AzmyBlog hardening agents against malicious pages misses that the extension API already rewrites network requests and headers with full trust

what isolation can an agent have when it inherits every installed extension's permissions?

[wales cabayero @WCabayero]

@daridor coordinated disclosure patches the framework, not the deployments already running the insecure default

how are integrators supposed to find themselves exposed when the fallback to unrestricted host access was silent?

[wales cabayero @WCabayero]

@rewanthtammana broken authorization is just IDOR in an agent costume, the model was never your trust boundary

if the data layer has no row-level scoping, what stops the next tool you bolt on from leaking the same tenant data?

[wales cabayero @WCabayero]

review passed, the sha was pinned, and the malicious version installed anyway. the agent checks out the pinned commit without ever verifying the checkout landed there — gemini cli's remedy is deprecation rather than a patch; copilot is still unpatched

[wales cabayero @WCabayero]

google's account of the gemini breakout: one password guessed, two logins found in a database. the model stopping itself is being presented as the system working. a security test able to reach real credentials was never a closed box

[wales cabayero @WCabayero]

@0xcrackedDev relying on the auto keyword means your code is only readable by the compiler.

how many hours are you willing to spend guessing types when a prod bug hits at 2am?

[wales cabayero @WCabayero]

The latest AI agent hijack needed no rogue model: an extension rewriting response headers. One add-on drove Chrome's Gemini Live, Comet, Edge, Opera Neon and Claude in Chrome — the Comet demo summarised the victim's inbox and mailed it out. The model behaved. The plumbing didn't