Code415 Radar

Key ideas

Goal drift under failure

An agent that keeps trying to finish a task can escalate from normal requests to attack patterns when the normal path is blocked.

SQL injection probe

A request that slips SQL syntax into a parameter to see whether the server builds queries from raw input.

Egress controls

Rules that limit which hosts and request types an agent's network traffic may reach.

Steps

1. Read the Transluce agent activity writeup first and note the exact sequence from normal fetch to probe.
2. Skim the SecurityWeek story for the university library and Data USA examples and the request counts.
3. Read the BleepingComputer piece on the Medicare portal to see how a research task crossed an access block.
4. Compare it with a web scraper you have written: where would a retry loop have stopped, and where would an agent keep going?
5. Self check: which single control, logging, egress allowlist or tool restriction, would have caught each incident earliest?

Steps

1. Start a small web app in a local container that returns errors on normal requests, so the agent's fetch keeps failing.
2. Give your agent a plain data fetching task against that local app only, with a fetch tool and no other network access.
3. Log every request the agent sends, including query strings and headers.
4. Scan the log for SQL keywords, path traversal patterns and script tags and count them per run.
5. Add a system prompt rule and a tool level allowlist, then rerun and compare the counts.

Small angles to try

• Run the same task on two model sizes and compare how quickly each escalates
• Add a retry budget to the tool and measure whether probing drops
• Replay the logs through a WAF rule set to see what it would have blocked

Key ideas

Kernel SRU

Stable Release Update, Ubuntu's process for shipping fixes to an already released version.

Proposed pocket

An Ubuntu archive area where updates wait for testing before reaching everyone.

CVE inflation

The upstream kernel now assigns CVE ids to almost any bug that could affect a running system, so counts rise fast.

Steps

1. Read Canonical's blog post on the new kernel release strategy and write down the old and new cycle lengths.
2. Read The Register article for the outside view on why the CVE count exploded.
3. Look at the two week cycle split: week one builds and boot checks, week two hardware certification.
4. Connect it to your own dependency updates: how often do you ship patches, and what testing slows you?
5. Self check: what risk do you accept if you pull kernels from the proposed pocket in production?

Steps

1. In a fresh Ubuntu VM, record the running kernel version and the installed kernel package versions.
2. Look up the Ubuntu security notices for your release and count kernel notices in the last 30 days.
3. Check whether Livepatch or unattended upgrades are enabled on the VM and note the settings.
4. Enable the proposed pocket only inside a throwaway VM, compare the candidate kernel version with the stable one, then disable it again.
5. Write a one page note: how many reboots per month would a weekly kernel mean for your servers?

Small angles to try

• Compare kernel notice counts for the same month in 2024 and 2026
• Script a check that reports how many days your hosts lag behind the newest kernel
• Repeat the count for another distro such as Debian or Fedora

Key ideas

Face swap

A model that replaces one person's face or appearance in video while keeping motion and lighting.

C2PA content credentials

A standard for signed metadata that records who made a file and which edits were applied.

Right of publicity

The legal idea that a person controls commercial use of their name and likeness.

Steps

1. Read the Futurism article for the timeline and Kalshi's response.
2. Read the C2PA explainer on the Content Authenticity Initiative site to see what a signed edit history looks like.
3. Look at how platforms show AI labels today and which edits trigger them.
4. Connect it to code signing: a signature proves origin, but only if the chain is kept when files are copied.
5. Self check: would C2PA have helped Choy if the agency simply stripped the metadata?

Steps

1. Take a photo with a phone or camera that writes content credentials, or export one from an editor that supports them.
2. Inspect the file with a C2PA verify tool and note the recorded author and edit steps.
3. Re-save the file through a messaging app or a screenshot and inspect it again.
4. Record which paths keep the credentials and which strip them.
5. Write down what a viewer could and could not prove about each copy.

Small angles to try

• Try the same with a short video clip
• Compare three social platforms' upload pipelines
• Add an invisible watermark tool and check whether it survives recompression

Key ideas

p75 latency

The time within which 75% of real user sessions finish, less noisy than the mean and fairer than p50.

Instruction counting

Measuring CPU instructions executed instead of seconds, so small wins are visible on any machine.

V8 code cache

Precompiled JavaScript bytecode saved to disk so an app skips recompiling at startup.

Steps

1. Read the claude.dev blog post from top to bottom and copy the before and after table.
2. Pick two fixes, the message tree lookup and the regex pre check, and write down why each saved work.
3. Look at how the rollout went from employees to 1% of users to everyone.
4. Compare with your own app: which of your pages has the worst p75, and do you even measure it?
5. Self check: why does instruction counting beat wall clock time when you ship thousands of small changes?

Steps

1. Pick one hot function in your own code, for example a parser or a regex heavy formatter.
2. Time it 20 times with wall clock timing and record how much the runs vary.
3. Measure the same function with an instruction count or a CPU profile in DevTools.
4. Add a cheap early check before the expensive regex, then measure again with both methods.
5. Record Largest Contentful Paint for your page in Lighthouse before and after one rendering fix.

Small angles to try

• Ask a coding agent to propose five micro fixes and keep only the ones that win by instruction count
• Test with non Latin text such as Japanese to see string handling costs
• Measure layout shift before and after keeping a component mounted

Key ideas

Zero day

A flaw attackers use before the vendor has shipped a fix.

HR system blast radius

HR platforms hold identity, family and address data, so one breach exposes far more than logins.

Unverified claim

A breach story where the data sample is checked but the entry path is not.

Steps

1. Read 404 Media's first report and note exactly what was verified.
2. Read BleepingComputer for the PeopleSoft angle and the claim about Fortune 500 companies.
3. Look up Oracle's Critical Patch Update schedule to see how often PeopleSoft gets fixes.
4. Compare with past ShinyHunters campaigns against SaaS tenants.
5. Self check: which facts in this story are confirmed, and which rest only on the group's word?

Steps

1. List the internet facing admin or HR apps your own team runs.
2. For each, find the vendor's last security advisory date and your deployed version.
3. Mark which ones sit directly on the internet and which are behind SSO or a VPN.
4. Estimate what personal data each would expose if fully read.
5. Rank them and write one mitigation per app.

Small angles to try

• Check whether your vendor publishes a patch calendar
• Map which apps share the same identity provider
• Repeat the list for a side project

Key ideas

Recursive resolver

The DNS server your ISP gives you that looks up names on your behalf.

Public DNS

Resolvers such as 8.8.8.8 or 1.1.1.1 that anyone can use instead of the ISP's.

Failure domain

The set of users who go down together when one component breaks.

Steps

1. Read the ITmedia article for the timeline and affected operators.
2. Read a basic explainer on how DNS resolution works, for example Cloudflare's learning center page on DNS.
3. Note what symptoms separate a DNS outage from a routing outage.
4. Connect it to your own apps: which ones hard depend on a single resolver or config service?
5. Self check: why did pinging an IP address still work while websites failed?

Steps

1. Look up the same domain with your default resolver, 8.8.8.8 and 1.1.1.1 and note response times.
2. Repeat 20 times each and record median and worst times.
3. Point your machine at an unreachable resolver on purpose and note which apps fail and how.
4. Configure two resolvers and confirm the fallback works.
5. Write down how long the fallback took to kick in.

Small angles to try

• Compare IPv4 and IPv6 lookups
• Try DNS over HTTPS in your browser
• Log resolver latency for a week with a small cron script

Key ideas

Plan mode

A Claude Code mode where the agent proposes a plan without editing files until you approve.

Reasoning effort

A setting for how much thinking a model does before answering, trading cost and speed for depth.

Human checkpoint

A point where a person reviews the agent's intent before it acts.

Steps

1. Read Thariq's post and a sample of the replies for the strongest arguments on both sides.
2. Read the Claude Code docs section on plan mode to see what it does today.
3. List the times plan mode saved you from a bad edit, or the times it only slowed you down.
4. Compare with code review: a plan is a design review before the diff.
5. Self check: if plan mode goes, what would you use instead to review intent before edits?

Steps

1. Pick one small refactor task in your own repo.
2. Run it once with plan mode and once without, from the same git commit.
3. Record time to finish, number of files touched and whether tests passed.
4. Repeat both runs at a lower and a higher effort level.
5. Compare which setting gave the best result per minute.

Small angles to try

• Use a bigger task that spans several modules
• Ask a teammate to judge the diffs blind
• Track token use for each run

Key ideas

Total ionizing dose

The accumulated radiation a chip absorbs over time, which slowly degrades it.

Single event upset

A bit flip caused by one charged particle hitting memory or logic.

Radiative cooling

In vacuum, heat can only leave by radiating from surfaces, so radiators must be large.

Steps

1. Read Google's Project Suncatcher facts post.
2. Read the DCD article for the long term plan of 81 satellite clusters.
3. Look up how ECC memory corrects single bit flips.
4. Connect it to data centers on Earth: power and cooling are the limits there too, just solved differently.
5. Self check: why is solar power attractive in orbit, and why is cooling harder there?

Steps

1. Train or load a small image classifier and record its accuracy.
2. Write a function that flips one random bit in a random weight.
3. Inject 1, 10, 100 and 1,000 flips and measure accuracy each time.
4. Repeat with the weights stored in float16 and in int8.
5. Plot accuracy against number of flips.

Small angles to try

• Flip bits only in exponent bits versus mantissa bits
• Add a simple checksum per layer and count how many flips it catches
• Try the same on a small language model's perplexity

Key ideas

Prompt prefix cache

Stored model state for the start of a prompt so repeated prefixes are cheaper and faster.

Cache breakpoint

A marker telling the API where a reusable prefix ends.

Hit rate

The share of input tokens served from cache.

Steps

1. Read OpenAI's post on better prompt caching for GPT-6.
2. Look at the API docs section on caching and find where cached tokens show up in the usage fields.
3. Compare with Anthropic's cache control, which also uses explicit breakpoints.
4. Think about your own prompts: which parts are static and which change per call?
5. Self check: why does putting the user message before the system instructions ruin the cache?

Steps

1. Build a prompt with a long fixed system section and a short changing question.
2. Send 10 requests and record cached token counts and latency from the usage data.
3. Move a changing value such as a timestamp to the top and repeat.
4. Add an explicit breakpoint after the fixed section and repeat.
5. Compute cost per request for each layout.

Small angles to try

• Change reasoning effort between turns and confirm the cache holds
• Wait past the reuse window and measure the first miss
• Run the same layout against another provider's cache

Key ideas

Harness

The loop and tools around a model that turn it into an agent: prompts, tool calls, retries, memory.

Terminal Bench

A benchmark of shell tasks used to compare coding agents.

Async tool dispatch

Running several tool calls at once instead of waiting for each.

Steps

1. Read the Strands Harness launch post and note which benchmarks and models were used.
2. Skim the Unreal Agent README for its design choices.
3. Read the HN thread for skeptical replies.
4. Compare with the agent loop you use most and list what it does differently.
5. Self check: why might a model score better inside its maker's own harness?

Steps

1. Pick five small shell tasks from your own work with a clear pass check.
2. Run them with your usual agent and with one of the two new harnesses on the same model.
3. Record pass rate, tokens used and wall time.
4. Repeat once to see how much results vary.
5. Write down cost per solved task for each.

Small angles to try

• Swap in a second model to see if the ranking flips
• Limit each harness to a single bash tool
• Count retries per task

Key ideas

NFC

Near Field Communication, short range radio used by contactless cards and phones.

EMV contactless

The card network standard that defines how a tap payment is authorized securely.

Tokenization

Replacing a real card number with a one time token so the device never stores it.

Steps

1. Read the Tom's Hardware article for the patent details.
2. Read a primer on how contactless EMV payments work, for example from EMVCo.
3. Look at how phone wallets use a secure element and tokens.
4. Connect it to in game stores: what friction does typing a card number add today?
5. Self check: what would Sony need, beyond an NFC chip, to accept card taps safely?

Steps

1. Scan an NFC sticker or transit card you own and note what data the app can read.
2. Tap your own contactless card and note what is and is not exposed.
3. Measure the maximum reading distance with a ruler.
4. Try reading through a phone case and a controller sized piece of plastic.
5. Write down which data a game console would actually need to take a payment.

Small angles to try

• Compare an iPhone and an Android phone
• Write a small NDEF record to a blank tag you own
• Time how long each tap takes

Key ideas

Consumer agent

An assistant that takes actions such as booking or form filling, not only chat.

Connector

An integration that lets an agent read or act inside another app.

Approval step

A prompt asking the user to confirm before the agent acts.

Steps

1. Read the TechCrunch article on the Charm device.
2. Open the App Store listing and read recent one star and five star reviews.
3. Note which tasks reviewers say it does well and where it fails.
4. Compare with the agent tools you use at work: what does a consumer version hide?
5. Self check: which permission would you refuse to give a phone agent, and why?

Steps

1. Install Muse and give it one low risk task, such as drafting a calendar event.
2. Record each approval prompt and what data it asked for.
3. Give the same task to another assistant app you use.
4. Compare steps, time and how clear the confirmations were.
5. Revoke the permissions afterwards and note how easy that was.

Small angles to try

• Try a task in Japanese
• Check what the privacy settings let you delete
• Time a booking flow against doing it by hand

Key ideas

Benchmark contamination

When test data appeared in training data, so scores measure memory rather than skill.

Semantics preserving rewrite

Changing names and layout of code without changing what it does.

Fault localization

Finding which file and line cause a bug before fixing it.

Steps

1. Read the paper abstract and method section on Hugging Face Papers.
2. Look at one SWE-bench Verified task to see what the model is given.
3. List the four kinds of rewrite the authors use.
4. Connect it to interviews: a candidate who memorized LeetCode fails on a reworded problem.
5. Self check: which rewrite would you expect to hurt most, and why?

Steps

1. Pick one fixed bug from the repo's history and write the issue text.
2. Ask your agent to fix it on the original repo and record steps and success.
3. Rename the package namespace and key functions with a refactoring tool, keeping tests green.
4. Ask the agent again with a reworded issue and record the result.
5. Compare steps spent on locating the bug.

Small angles to try

• Try a private repo the model cannot have seen
• Only reorder files without renaming
• Compare a small and a large model

Key ideas

Output escaping

Encoding user values so they are treated as data, not markup or code.

OG image route

A server route that renders social preview images, often from URL parameters.

Incoming email token

A secret address that lets email create issues or merge requests as you.

Steps

1. Read The Hacker News article on the Next.js flaw and note the affected versions.
2. Read the Next.js release notes for 16.3.6.
3. Read the GitLab email token article and the reason HackerOne closed it.
4. Connect it to XSS: the same missing escape, but on the server.
5. Self check: which of your routes pass query parameters into rendered images or templates?

Steps

1. Check the installed Next.js and Satori versions in your lockfile.
2. Search your code for ImageResponse and note which inputs reach it.
3. Upgrade to a fixed version in a branch and run your tests.
4. In GitLab, find where your incoming email token is shown and reset it.
5. Write a checklist item for future OG image routes.

Small angles to try

• Add a dependency alert for Satori
• Check whether your routes run on Node or Edge
• Audit other template renderers for raw input

Key ideas

Ternary weights

Each weight is one of three values, so it needs under two bits.

Hadamard rotation

A fixed orthogonal transform that spreads outliers so low bit rounding loses less.

GGUF

The file format used by llama.cpp and many local runners.

Steps

1. Read the Hugging Face model card and its benchmark table.
2. Read a short explainer on BitNet style ternary models.
3. Compare sizes: FP16, 4 bit and this 1.75 bit file.
4. Connect it to JPEG: lossy compression that keeps what matters most.
5. Self check: why does rotating weights before rounding reduce error?

Steps

1. Download the 1.75 bit GGUF file from the model page.
2. Load it in your GGUF runner and record load time and memory use.
3. Ask 10 questions from your own work and save the answers.
4. Run the same questions on a 4 bit 8B model you already use.
5. Compare quality, tokens per second and memory.

Small angles to try

• Try the 2.13 bit version too
• Test Japanese prompts
• Measure perplexity on a text file of your own

Key ideas

XML signature wrapping

Moving signed XML elements so a verifier checks one element but the app reads another.

Parser differential

Two parsers reading the same input differently, which attackers can exploit.

OIDC

OpenID Connect, an identity layer on OAuth 2 that uses signed JSON tokens.

Steps

1. Read the Trail of Bits post.
2. Skim a diagram of a SAML login flow and an OIDC login flow.
3. Note where each one verifies signatures.
4. Connect it to JSON vs XML parsing in your own code.
5. Self check: why is an enveloped signature harder to verify safely than a detached one?

Steps

1. Run a local identity provider in a container.
2. Set up one test app with SAML and one with OIDC against it.
3. Log in with both and capture the assertion and the ID token.
4. Compare size, fields and where the signature sits.
5. List which libraries in your stack handle each.

Small angles to try

• Check which SAML library version your company uses
• Decode the ID token and verify it by hand with the public key
• Time both login flows

Key ideas

Scheduled capture

The app prompts or records at set times instead of when the user chooses.

Background camera limits

iOS restricts camera use when an app is not in the foreground.

Clip stitching

Joining short video segments into one file with transitions.

Steps

1. Open the App Store page and read the description and recent reviews.
2. Read the coelia.jp explainer for the privacy concerns.
3. Look up how iOS notifies users when the camera is on.
4. Connect it to BeReal and other time based social apps.
5. Self check: why must the app nudge you each hour instead of filming silently?

Steps

1. Record six 2 second clips with your phone across one afternoon.
2. Copy them to your laptop.
3. Join them into one video with ffmpeg and note the file size.
4. Add a timestamp overlay to each clip.
5. Compare the result with what setlog produces.

Small angles to try

• Add a short crossfade between clips
• Automate it with a small script
• Measure how much battery hourly recording would use

Key ideas

Browser extension bridge

An extension that receives commands from a local daemon and acts in the browser.

Session reuse

The agent uses your existing cookies, so it is logged in wherever you are.

Git worktree

A second working folder for the same repo, useful for parallel experiments.

Steps

1. Read the BrowserSkill README and its architecture section.
2. Note what permissions the extension asks for.
3. Skim the OpenResearch README for how it runs parallel sessions.
4. Compare with Playwright, which starts a fresh browser.
5. Self check: what is the risk of an agent that browses with your logged in cookies?

Steps

1. Create a new browser profile with no personal logins.
2. Install BrowserSkill following its README into that profile only.
3. Ask your agent to open a public page and extract a table.
4. Record each command it sent and how long it took.
5. Uninstall and remove the profile when done.

Small angles to try

• Compare with a Playwright script for the same task
• Try a site that needs scrolling
• Check what the daemon logs

Key ideas

Alpha channel

A fourth image channel storing transparency for each pixel.

Reference images

Extra inputs that guide identity, style or layout.

Masked edit

Changing only a selected region while keeping the rest.

Steps

1. Read the model card on Hugging Face.
2. Read the license terms and note what commercial use needs.
3. Look at the sample outputs for text rendering.
4. Compare with your current icon workflow and background removal.
5. Self check: why is native transparency better than cutting out a background later?

Steps

1. Generate three sticker style images with transparent backgrounds.
2. Open them in an editor and check the alpha edges.
3. Edit one image with a mask and one reference image.
4. Render a short Japanese and an English text label.
5. Compare with the same prompts in a model you already use.

Small angles to try

• Use 5 references instead of 1
• Check how edges look on dark and light backgrounds
• Time generation per image

Key ideas

Accelerator tier

The class of GPU or TPU a notebook runtime gets.

Background execution

The notebook keeps running after you close the tab.

Compute units

Colab's usage currency for paid runtimes.

Steps

1. Read the Google Developers Blog post.
2. Open Colab and look at the runtime type menu.
3. Check which plan your account has.
4. Compare with renting a GPU by the hour elsewhere.
5. Self check: what job would you only run if background execution works?

Steps

1. Open a new notebook and check which accelerators are offered.
2. Run a small benchmark such as a fixed matrix multiply loop and record time.
3. Switch to another accelerator if offered and repeat.
4. Record the GPU name the runtime reports.
5. Note the result in a table by plan.

Small angles to try

• Train a tiny model for 10 minutes and compare cost
• Close the tab and see whether the job continues
• Compare with your local GPU