Security & Trust

Upload a photo and break an image classifier on purpose. Craft subtle FGSM or PGD perturbations, a visible adversarial patch, or a black-box attack with no gradient access, untargeted or aimed at a specific label. Then try two inference-time defences, check whether the attack transfers to a second model, and see adversarial training compared against a standard model on the run you just performed. It reports honestly whether a defence actually recovered the right label, and whether a targeted black-box attack converged at all within the query budget — often it doesn't.

Paste a code snippet and see the stylometric signals people associate with AI authorship — comment density, generic naming, docstring formality, exception handling, boilerplate phrasing — alongside an independent LLM opinion, shown side by side. It deliberately never returns a probability or an 'AI-written' verdict, because no reliable general-purpose detector exists in the published research and a confidence number here would be invented.

Enter a domain and see what it exposes to the open internet. Four passive checks run live: sensitive paths like .git/HEAD and .env (only flagged when the response really is that file, not merely a 200), Apache/nginx directory listings, CMS fingerprinting from the standard generator tag, and a short common-port connect check. It refuses to touch private, loopback or internal addresses, and reports real findings for you to weigh rather than a made-up risk score.

Upload any file and see its structure as a picture. The bytes are rendered as the grayscale byte-plot used in malware-visualisation research, next to a sliding-window entropy heatmap — sustained near-random entropy is an established sign of packed or encrypted content, the same signal tools like PEiD look for. Windows executables also get a PE header check for a classic packer tell. It won't name a malware family — no dependable pretrained model exists for that — and it never executes the file: static byte analysis only, up to 5MB.

Paste a Chrome or Edge extension's manifest.json and see what it is allowed to do. Checks individually-risky permissions (debugger, nativeMessaging, webRequestBlocking, cookies, history), broad host access, and dangerous combinations — broad host access plus network interception plus cookies together enable session hijacking on any site. This reads declared permissions, not behaviour: a legitimate password manager needs much the same access, so findings are framed as worth a closer look, never a judgement of intent.

Upload a CAPTCHA-style image and watch a vision-language model try to read it — modern VLMs handle plain text CAPTCHAs far more easily than classic OCR ever did. One intensity slider then stacks three model-agnostic hardening techniques (pixel noise, an occlusion wave, contrast reduction) and the model tries again, side by side. Nothing gradient-based is used, because the solver here is a black box — the same constraint a real CAPTCHA vendor faces. It only ever reads an image you upload; it never contacts a live CAPTCHA on a real site.

Paste a DNS query log, or check a single hostname, and spot possible tunnelling or exfiltration. Uses the published heuristics real tools rely on for this (MITRE ATT&CK T1071.004): subdomain length, Shannon entropy and query volume per parent domain. A domain is only flagged when several signals agree, so ordinary long CDN-style subdomains don't trip it. Pure heuristics, no model, fully client-side.

Paste raw email headers and see whether the sender checks out. You get two things: what the receiving mail server's own Authentication-Results already concluded about SPF, DKIM and DMARC (relayed, not re-verified), and independent live DNS lookups of the sending domain's real records, plus a From: alignment check. It does not cryptographically verify the DKIM signature — that needs the full message body — and says so rather than implying otherwise.

Add a barely-visible perturbation to a photo so face-recognition models place it somewhere other than your real face. A simplified take on Fawkes, the privacy technique built to counter unauthorised facial-recognition scraping. You get the actual measured drop in embedding similarity, and an honest caveat: this protects the copy you cloak, not photos of you already scraped elsewhere.

See how face re-identification actually works, on photos you supply. Upload a target photo and a small gallery, and the gallery is ranked by how closely each face matches — a real measured similarity, the same mechanism behind Clearview-style search. A 'Protect and re-test' step then cloaks the target and runs the identical search again so you can see whether the match survives. It searches nothing but the photos in your request — no internet, no database.

Type a short phrase three times to enrol a keystroke-timing profile, then type it once more and see how closely the rhythm matches. Scoring uses scaled Manhattan distance over dwell and flight times, a published approach for keystroke-dynamics anomaly detection. Try typing normally, then deliberately faster or hunt-and-peck, and watch the score move. A concept demo rather than a calibrated authenticator — and entirely client-side, with no server call.

Paste a prompt, or a document an AI might be asked to read, and see whether it tries to hijack the model. Two independent signals sit side by side: a transparent pattern library covering direct overrides, jailbreak roleplay, indirect injection and encoding tricks, and a separately-prompted LLM judge. They combine into an overall risk badge rather than one invented confidence number — no detector here is claimed to be reliable on its own.

Paste a package.json, requirements.txt or a source file and see what a supply-chain reviewer would flag. Checks for npm install-script hooks, dependency names that typosquat well-known packages, dynamic execution calls (eval, exec, subprocess), obfuscated high-entropy strings, embedded URLs, hardcoded secrets, SQL built by string interpolation, and unsafe deserialization. It matches attacker techniques rather than known signatures, which is what lets it flag packages nobody has seen before. Every hit is real evidence to judge, never a safe/malicious verdict. Runs fully in your browser.

Check how strong a password really is. Scored in your browser by zxcvbn, the pattern-matching algorithm behind many real password meters — dictionaries, keyboard walks, dates, repeats — rather than naive character-class counting. You can also check it against Have I Been Pwned using k-anonymity: only the first five characters of its SHA-1 hash ever leave your machine, never the password itself. Nothing is stored.

Paste an email's body text and see whether the writing itself reads like phishing — urgency, generic greetings, manipulative phrasing. A Multinomial Naive Bayes classifier trained on real phishing and legitimate mail shows you the exact words driving its score, next to a separate, transparent list of rule-based flags. Two signals shown side by side, never blended into one black-box number. Runs fully client-side — nothing you paste leaves your browser.

Upload a photo or screenshot of a QR code and see where it actually points before you trust it. The decoded URL is checked for structural phishing signals — IP-literal hosts, punycode, '@' auth tricks, shorteners, suspicious TLDs, and typosquats of well-known brands by edit distance. The link is decoded and read, never visited. You get flags to weigh, not a binary safe/malicious answer.

Paste raw alert lines and get them grouped and prioritised. Near-identical alerts are deduplicated by template in your browser first, so only the grouped summary — never your raw log — is sent on to an LLM for a priority, a one-line reason and a suggested next step per group. Advisory only: every suggestion is written for you to act on, never phrased as something already done.

Add a barely-visible perturbation across an image so its CLIP embedding drifts away from where a model would naturally place it — a simplified take on the Glaze and Nightshade approach to countering AI style-mimicry. You get the actual measured similarity drop, calibrated against an unrelated-image baseline, plus the honest caveat: it protects the copy you cloak, not images already scraped elsewhere.

Enter a domain and check its TLS and security headers the way Mozilla Observatory does. A real handshake verifies the certificate chain, expiry and protocol version, flagging deprecated SSLv3 and TLS 1.0/1.1, and a live request checks the six standard security headers. It refuses to connect to private, loopback or internal addresses, and gives a qualitative verdict with the actual warnings behind it rather than a numeric score.

Upload a short clip of someone typing and recover when the keys were pressed from hand motion alone. Frame-by-frame hand tracking feeds a tap detector on fingertip movement, producing a timeline of keystrokes, which hand, and likely word boundaries from the gaps — the same side channel behind published research on video keystroke inference. It stops at timing and does not attempt to recover what was typed: that needs per-target trained models this doesn't have. Runs in your browser; no video leaves your device.

Scan a file with real YARA — the same pattern-matching engine antivirus and threat-intel teams use to write and share detection rules. Run it against a small built-in rule set (EICAR, PowerShell LOLBin encoding, webshell and macro patterns, embedded-PE smuggling, an entropy rule), or write your own rule and test it, which is what YARA actually exists for. Your file is never executed, and every hit shows the matched string and offset rather than a bare verdict.