Dylangzhan

AI agent engineer. USC Viterbi MS CS.

[Dylangzhan @dylangzhan]

Grok Bot can now call businesses for you, sit on hold, and press 1 for English.

The real AGI benchmark was never the bar exam. It was surviving a 45-minute hold with the cable company.

[Dylangzhan @dylangzhan]

the jeep smog story beats most agent benchmarks.

1994 cherokee, no OBD2, few docs. cheap arduino on the diagnostic plug, then claude and codex reverse-engineered the protocol, logged real drives, found the lean fuel trim.

it wasn't guessing. it had live data to check against.

[Dylangzhan @dylangzhan]

railway's postmortem says the quiet part: AI provider keys are now the most valuable secret in your env. the attacker decrypted vars across up to 100k workspaces hunting for them.

if your agents hold a raw key: spend cap per key, one key per service, alert on usage spikes.

[Cyber Security News @The_Cyber_News]

‼️ Railway Reveals How Hackers Used AI Agents to Steal API Keys and Rack Up $740,000 in Charges

Read More: https://t.co/v0GfdO2WnI

Railway has revealed how an attacker used AI agents, a build system flaw, and cloud access mistakes to steal customer secrets and generate an estimated $740,000 in unauthorized charges.

Its November 10 report describes an attack that crossed several internal systems while engineers worked to close multiple entry points. The findings expand on the earlier Railway security incident disclosure.

Up to 1.85% of workspaces, roughly 100,000, could have had environment variables decrypted. This is an upper limit, not a confirmed count of workspaces with stolen keys, because available logs identify affected environments rather than individual secrets.

#CybersecurityNews

[Dylangzhan @dylangzhan]

@AGTPinsights the takeaway for anyone running builds with secrets: assume the build box sees everything, and rotate whatever was ever exposed to it, not just what you think leaked

[Dylangzhan @dylangzhan]

@NovaIAHQ classic lead-agent failure. worker returns the fix, lead summarizes it and moves on without applying it. making the lead point at the merged diff before it can mark done is a cheap check

[Dylangzhan @dylangzhan]

@naumowf only if each retry gets something new, like the failing test output. same context in, same mistake out

[Dylangzhan @dylangzhan]

@AdamJWenborn the per-task cost gap is the real story. 0.4 points over fable max at under a third of the price means most teams should be routing, not picking one model

[Dylangzhan @dylangzhan]

@michaeldeayaw on error is easier to trust. a cheap pass plus a check that can actually fail (tests, schema) beats guessing difficulty before the first call

[Dylangzhan @dylangzhan]

google basically shipped the routing advice with the model: argon for most work, carbon for your hardest problems.

so don't make carbon the default in your agent loop. hand it the long debug or the step that already failed twice, keep a cheaper model on routine tool calls.

[Google AI @GoogleAI]

Our strongest frontier model, Gemini 4 Carbon, is now available.

Built for deeply complex, long-horizon workflows that push a frontier model’s intelligence, Carbon delivers frontier performance in agentic workflows across complex knowledge work, software engineering and web development, cybersecurity, and scientific reasoning.

While Gemini 4 Argon remains the ideal model for the majority of software engineering and knowledge work, Carbon steps in for your hardest tasks that push a model to its limits — from complex, multi-step debugging to open-ended scientific research.

Available starting today:
— Consumers: use it in the @GeminiApp and AI Mode in Search for Google AI Ultra subscribers
— Developers and Enterprises: build via the API in @GoogleAIStudio and @AndroidStudio, Google @Antigravity for Google AI Ultra subscribers, and Gemini for business
— Scientists and Researchers: apply for a chance to get 6 months free via @Antigravity
— Cybersecurity Defenders: apply to join the Fairwind Program for access to Carbon for cyber

Watch how Carbon brings reasoning and coding together, turning avian research into an interactive 3D simulation of a songbird’s vocal organ that synthesizes the sounds of six bird species live.

[Dylangzhan @dylangzhan]

@yhiyo2011 makes sense. terminal-bench rewards getting a command to work, a mergeable pr has to fit the codebase, pass tests and survive review

[Dylangzhan @dylangzhan]

@samongaro_ how much code you delete afterward is an underrated metric. diff size vs task size would make a decent eval

[Dylangzhan @dylangzhan]

@Falconiererb yeah, and a hook failing is loud, which is the point. a skipped skill just looks like the model being lazy

[Dylangzhan @dylangzhan]

@DrivenByNikiG terminal-bench is the number i'd watch. 5+ points over opus on agentic tasks at $2/$10 is a routing change if it holds on real repos

[Dylangzhan @dylangzhan]

@saiitoshii only with a hard spend cap on the key itself. approval on deletes is nice but the expensive mistake is a create stuck in a retry loop

[Dylangzhan @dylangzhan]

opus 5.5 vs fable 5.2 is a routing problem, not a loyalty contest.

if the $200 plan buys ~5x more opus than fable, fable shouldn't be running your agent loop. let opus do the edits and tool calls, and escalate to fable only for planning or a step that already failed twice.

[Theo - t3.gg @theo]

The $200 Claude Ultra plan gets you $15,000 of Opus 5.5 in a month, but only around $3,000 of Fable 5.5.

Even ignoring that Fable is restricted to 50% of your weekly limit, it burns faster per $ than Opus does.

More info below

[Dylangzhan @dylangzhan]

@xinwei_97 logging the commit sha with every trace fixes half of this. replay against that exact tree, not main

[Dylangzhan @dylangzhan]

@dng_notes exactly. diff the test files before you trust the green, most of the cheating shows up there

[Dylangzhan @dylangzhan]

@ganesh2412 yeah, review flag not defect label. if it cries wolf on every diff people stop reading it fast

[Dylangzhan @dylangzhan]

@saiitoshii good catch. ForwardAgent yes under a Host * block is the scary version, every host the agent touches can sign as you