Ai security icons

The 52 ai-security icons in Security, described

Adversarial exampleadversarial-example

An adversarial example — looks the same to a person but classifies as something else.

Backdoorbackdoor-model

A backdoor — a hidden keyhole in a model that nobody was told about, a trojan.

Canary tokencanary-token

A canary token — a decoy credential that raises the alarm the moment it is touched.

Capability evalcapability-eval

A capability eval — measure what a model can do, and how dangerous, before release.

Circuit breakercircuit-break

A circuit breaker — the wire opens before the damage does, calls cut off to fail fast.

Content credentialcontent-credential

A content credential — the page carries its own birth certificate, C2PA provenance.

Data exfiltrationdata-exfil

Data exfiltration — data that left through the wall, a leak or a breach.

Deepfake detectdeepfake-detect

Deepfake detection — spotting the face that never had a body, a synthetic image caught.

Detect AI textdetect-ai-text

Detect AI text — a classifier reading for the machine's accent.

Dual usedual-use

Dual use — a capability that cures with one edge and cuts with the other.

Evasionevasion

Evasion — an attack that slips straight through the defence without being caught.

Export controlexport-control

Export control — a model or a chip that may not cross the border, restricted.

Harm categoryharm-category

A harm category — which kind of bad this content is, a class in the safety taxonomy.

Hazardhazard

An alert mark inside a diamond — a dangerous capability flagged before release.

Honeypothoneypot-ai

A honeypot — bait that knows your name, a decoy that traps an attacker.

Sanitise inputinput-sanitize

Sanitise input — scrub what comes in before the model sees it.

Interpretabilityinterpretability

Interpretability — the lens finally goes inside the model to explain it.

Jailbreakjailbreak

A jailbreak — out through the broken bar, a prompt that bypasses the guardrails.

Kill switchkill-switch

A kill switch — the one emergency control that ends it right now.

LLM firewallllm-firewall

An LLM firewall — the guard that decides what a model is not allowed to say.

Membership inferencemembership-inference

Membership inference — an attack that tells whether a record was in the training set.

Misusemisuse

Misuse — the right tool in the wrong hands, abuse of a capability.

Model poisoningmodel-poisoning

Model poisoning — tainted data slipped into training on purpose.

Model theftmodel-theft

Model theft — the whole model extracted and walked out the door.

Sanitise outputoutput-sanitize

Sanitise output — scrub a response clean before anyone sees it.

Oversightoversight

Oversight — somebody is actually watching the system as it runs.

PII redactionpii-redact

PII redaction — the personal part blacked out so privacy is kept.

Policy allowpolicy-allow

Policy allow — the rules read yes, the request may go ahead as asked.

Policy blockpolicy-block

Policy block — the rules read no, the request is stopped before it runs.

Safety probeprobe-safety

A safety probe — a thin question lowered into the model to test what it knows.

Prompt injectionprompt-injection

Prompt injection — malicious instructions smuggled into a model's input to hijack it.

Prompt shieldprompt-shield

A prompt shield — the prompt goes into the model guarded against injection.

Provenance chainprovenance-chain

A provenance chain — every hand something passed through, linked in order.

Red teamred-team

Red team — attacking your own system before somebody else does, adversarial probing.

Redact fieldsredact-fields

Redact fields — the line is there but the words are hidden behind a bar.

Run scanrun-scan

A shield with a play inside — start a security scan right now.

Safety filtersafety-filter

A shield with a funnel inside — the filter that stops unsafe content.

Sandbaggingsandbag

Sandbagging — a model stronger than it lets on, hiding capability.

Scan stoppedscan-stopped

A shield with a stop square inside — a security scan halted before it finished.

Secret scansecret-scan

Secret scan — find the credential somebody committed to the code.

Security trendsecurity-trend

A shield with a rising line inside — how the security posture is trending.

Shield upgradeshield-upgrade

A shield, a double chevron rising inside — protection raised to a higher level.

Signed mediasignature-media

A picture with a signature beneath it — media signed to prove where it came from.

Sleepersleeper

A sleeper — a model that behaves until the trigger phrase wakes its hidden behaviour.

Threat eventthreat-event

A shield beside a lightning bolt — a security event that hit the guardrails.

Threat huntthreat-hunt

A shield beside a magnifying glass — searching for threats before they strike.

Trigger phrasetrigger-phrase

A trigger phrase — say the words and the model's behaviour changes.

Tripwiretripwire

A tripwire — cross the line and it is known, a detection that fires.

Virtual patchvirtual-patch

A virtual patch — armour over the hole until the real fix lands.

Watermarkwatermark

A watermark — an invisible stamp that marks content as generated, proof of provenance.

Watermarkwatermark-ai

An AI watermark — content signed invisibly in the words themselves for provenance.

Weight leakweight-leak

A weight leak — the model's weights dripping out of the building, theft or exposure.

In code, each is one import — import { AdversarialExample } from "@iconmind/react/icons/adversarial-example" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag