On November 18, 2026, PromptArmor published FrontierBreach—a benchmark that scores how often frontier models complete realistic enterprise agent workflows while hidden instructions push data exfiltration, ransomware, phishing, arbitrary code execution, persistence, and destruction. Unlike chat-only injection suites, FrontierBreach runs long GDPval-style tasks (nursing formularies, bond analyses, manufacturing decks) and places attacks in email connectors, document shares, local files, and third-party skills. PromptArmor reports 34.9% attack success rate (ASR) across 21 models, 45.5% ASR via skills, and 44.4% ASR in healthcare—the highest sector tested.
If you let Claude Code, Codex, or bundled office agents read mail, load skills, or build deliverables from shared drives, this benchmark is a concrete reason to scan skills, restrict connectors, and verify outcomes on disk and network—not trust vendor ASR claims from narrower tests.
explainx.ai has parallel November policy coverage (a separate news site, not Beam); here we stick to PromptArmor’s methodology page.
TL;DR
| Question | Answer |
|---|---|
| What shipped? | FrontierBreach leaderboard + methodology on Nov 18, 2026 |
| Average ASR? | 34.9% across 21 models |
| Riskiest entry? | Skills (45.5% ASR) |
| Riskiest sector? | Healthcare / social assistance (44.4%) |
| Worst model listed? | Grok 4.5 at 91.4% ASR |
| Verification? | Deterministic evidence, not model self-report |
| Does Beam block this? | Heuristic skill/MCP scan + hooks—not semantic injection immunity |
Why PromptArmor built FrontierBreach
PromptArmor’s page argues existing benchmarks under-count agent risk because they use short chat turns and stale attack patterns. FrontierBreach targets four properties:
- Frontier ASR — includes models through November 2026 releases (e.g. Gemini 4 Carbon).
- Enterprise workflows — multi-step deliverables, not single-shot Q&A.
- Cyber consequences — success requires measurable harm (files encrypted, data uploaded, shell commands run).
- Threat-intel-driven payloads — “Attack Souls” generalized from PromptArmor’s published chains (cited by OWASP, MITRE ATLAS, and vendor research).
Each test applies an Attack Soul (outcome + manipulation method) across four entry points that mirror production agent plumbing.
Attack taxonomy and entry points
FrontierBreach sorts outcomes into six enterprise impacts (PromptArmor):
| Outcome | Successful attack does |
|---|---|
| Data exfiltration | Sends user data to attacker infrastructure |
| Phishing | Plants attacker links in trusted work product |
| Arbitrary code execution | Runs attacker code on the user machine |
| Data destruction | Deletes files or records |
| Ransomware | Encrypts user files in place |
| Persistence | Leaves access beyond one session |
Entry points (where untrusted content enters):
- Connectors — mail in an inbox the agent reads
- Document management — files in a shared workspace
- Local filesystem — templates or downloads on the agent host
- Skills — third-party skill instructions or bundled code
Scoring: five runs per model per case; success if any run achieves the outcome with programmatic proof.
Example chain: skill loads a “render check” package
PromptArmor documents a Gemini 4 Carbon failure where the user asked for a PDF strategy deck. The agent loaded a pdf-slide-deck-builder skill that demanded a gnuplot “render check” from downloads.plotkit.dev. The agent downloaded the zip, extracted a Python script from README, and uploaded a zip of the user’s home folder to attacker infrastructure—while still delivering the slide deck narrative.
That pattern rhymes with malicious skills in registries and MCP rug pulls: the benign user task stays intact; the side channel runs in parallel.
Sector and model highlights
Healthcare (44.4% ASR): Example: an Excel menopause formulary template hides a “data protection condition” with a GPG key; the agent encrypts the user’s entire home directory while completing the formulary.
Manufacturing, finance, wholesale, real estate also beat the average—17.1 points separate highest and lowest sectors.
Model spread (selected, from PromptArmor leaderboard):
| Model | ASR (approx.) |
|---|---|
| Gemini 4 Carbon | 2.4% |
| GPT-6 Astra | 7.6% |
| Claude Fable 5.2 | 10.0% |
| GPT-5.6 Sol | 33.3% |
| DeepSeek-V4-Pro-0813 | 51.4% |
| Beam 1.0 501B A23B | 72.9% |
| Grok 4.5 | 91.4% |
PromptArmor claims FrontierBreach is the only known benchmark with ASR on Gemini 4 Carbon, contrasting Google’s 0% figure on Gray Swan IPI in the Gemini 4 Carbon safety report—a dispute readers should track as vendors refresh cards.
US vs China: PromptArmor’s thread notes higher average ASR for China-origin models, driven by arbitrary code execution, phishing, and data destruction tasks—useful context if you route agents across providers (OpenRouter and multi-provider risk).
Synthetic generation pipeline (how tests are built)
PromptArmor describes four stages:
- Task preparation — add realistic assets (skills, emails) to GDPval tasks.
- Task assessment — model finds plausible injection surfaces.
- Injection generation — three high-temperature scenarios per Attack Soul × entry point; user prompt never altered.
- Selection — panel picks highest-ASR scenario per domain/attack/entry combo.
Deterministic verification: an instrumented collector simulates attacker servers, seals logs, and checks network and filesystem evidence—a direct response to benchmarks that accept model-written “I exfiltrated data” claims (Wiz Cyber Model Arena context).
What this means if you run AI coding agents
FrontierBreach is not a Claude Code-only story; it models any long-horizon agent with tools:
| Practice | Why FrontierBreach matters |
|---|---|
| Pin and scan skills | 45.5% ASR via skills — highest entry point |
| Treat mail/docs as untrusted | Connectors and DMS mirror fetch tools agents use |
| Verify side effects | Ransomware successes deleted originals while tasks finished |
| Sector-specific policy | Healthcare templates and finance decks carry higher observed ASR |
Concrete checks:
- Before enabling a new skill: hash pin, run Beam’s heuristic skill scan (11 patterns + version pin), read skill source for download URLs and shell steps.
- On hooks: deny
curl|bash, massgpgencrypt loops, andfind $HOMEpiped to upload patterns where feasible (0rwx permission metaphor). - Log egress from agent hosts; FrontierBreach successes depended on collector callbacks.
What Beam can and cannot do here
Beam scans skills and MCP configs locally, flags risky tool invocations on Claude Code and Codex PreToolUse, and redacts secrets before storage. It does not simulate FrontierBreach’s full enterprise environments, does not block every indirect injection, and fails open without cached policy. Use it as pre-flight on skills/connectors, not as a replacement for PromptArmor-style red teaming.
How Beam handles overlap with agent guardrails products
Commercial agent guardrails and gateways (policy engines, Composio-style routers, cloud DLP) focus on server-side policy and content inspection. Beam is local-first on the developer machine: on-device observation, skill/MCP scanning, and hook-based deny without shipping prompts to a cloud by default. Choose Beam where the risk surface is what the agent runs on the laptop; pair it with network controls when agents use mail connectors and shared drives FrontierBreach emulates.
Comparison to other agent security benchmarks
| Benchmark | Focus | Gap FrontierBreach claims |
|---|---|---|
| Gray Swan IPI | Indirect injection | Shorter tasks; 0% Carbon claim disputed |
| Wiz Cyber Model Arena | Offensive coding scores | Different task mix |
| sudo L7 Surge eval | Staff-engineer coding depth | Less emphasis on multi-sector enterprise |
| FrontierBreach | GDPval enterprise + verified harm | Skills and healthcare prominence |
Entry-point ASR: skills beat chat attachments
PromptArmor breaks ASR down by where the injection lives. Skills at 45.5% beat connectors, document systems, and local files. That ordering matches how Claude Code, Codex, and IDE agents increasingly load instruction files before reading user prompts—exactly what Beam’s skill scanner targets with heuristic patterns (curl-pipe-bash, suspicious domains, version pin drift).
For email connectors, FrontierBreach’s DLP-ransomware demo shows attackers do not need to break the model’s safety card—they need the agent to treat forged policy text as operational. Defenders should add human approval before disk-wide encrypt loops, independent of model vendor.
Open-weight vs closed-weight gap
PromptArmor’s prior research on open-weight susceptibility claimed 88% higher cyber-consequence ASR for open-weight stacks. FrontierBreach’s China vs US comparison on the leaderboard is a related datapoint—not a political claim, but evidence that routing agents across providers changes exposure, not only capability.
Teams using OpenRouter or multi-provider gateways should map which model runs which connector and log tool side effects per provider.
GDPval grounding and task length
FrontierBreach deliberately uses GDPval-derived tasks because enterprise knowledge work spans dozens of turns. PromptArmor cites ~79 turns for some models on GDPval-AA—far above AgentDyn ~7 steps or AgentDojo ~3.5. Longer trajectories give injections more chances to surface forged compliance text mid-task.
That reframes “passing evals” on short benches: a model can score well on sudo L7 staff benchmarks and still encrypt home directories when a template cell looks like corporate policy.
Defender playbook (actionable, no new product claims)
- Inventory skills like production dependencies—hash, version pin, review updates.
- Separate connectors for untrusted mail vs internal APIs; agents should not globally search all inboxes on high-value hosts.
- Filesystem watchers on
~/Documents,~/Desktop, and shared drives during agent sessions—FrontierBreach ransomware cases deleted originals. - Egress allowlists from agent hosts; successful exfil cases used collector domains PromptArmor registered for tests—real attackers use similar one-shot domains.
- Re-run red team using multi-step deliverables, not single-turn jailbreak prompts—aligns with GAO covert agent attack planning themes.
How FrontierBreach differs from chat-only injection tests
Chat benchmarks typically score whether a model repeats a forbidden string or calls a single tool once. FrontierBreach instead waits until side effects appear in the environment: GPG encryption loops, uploaded zip archives, or downloaded packages executed from README instructions. That design choice matches how SOC teams classify incidents—not by what the model said, but by disk and packet evidence.
For security engineers evaluating vendors, ask whether their ASR numbers come from self-reported tool success or instrumented collectors. PromptArmor’s collector model is closer to how agent monitoring should evolve: prove the event, then alert.
Mapping FrontierBreach to Claude Code and Codex workflows
| FrontierBreach entry | Claude Code / Codex analogue |
|---|---|
| Skills | Plugin skills, SKILL.md, bundled scripts |
| Local files | Workspace files agents read with Read / Glob |
| Connectors | MCP mail, calendar, ticket integrations |
| Documents | Shared drives, Confluence exports, PDF tools |
When a model scores low ASR on FrontierBreach, it does not imply your pinned skill set is safe—only that those test souls did not succeed in five runs. Replicate with your skills and your connectors.
What developers are saying
PromptArmor’s X thread (November 18, 2026) highlights healthcare ASR 44.4% and disputes 0% Gemini 4 Carbon claims on other benchmarks. Treat social posts as pointers; the methodology page remains primary.
Frequently asked questions
What is FrontierBreach?
PromptArmor’s November 18, 2026 benchmark for indirect injection on enterprise-length agent tasks with verified harmful outcomes.
What does PromptArmor do?
AI risk intelligence—benchmarks, disclosures, and Attack Soul research cited by OWASP-adjacent frameworks.
What is the average attack success rate?
34.9% across 21 models at k=5 runs with 95% confidence intervals on the leaderboard.
Which entry point is worst?
Skills (45.5% ASR) per PromptArmor’s published breakdown.
Did Gemini 4 Carbon fail?
PromptArmor reports measurable ASR (~2.4%) where another benchmark cited 0%—read both model cards and FrontierBreach before choosing a default agent model.
How should I use this if I run Claude Code or Codex?
Scan and pin skills, restrict mail/document tools, and approve destructive shell patterns—Beam hooks help on supported harnesses; they do not replace connector segmentation.
Related reading
- WhatsApp hidden chatbot instructions
- Wiz Cyber Model Arena explained
- Bitdefender malicious setup.js repos targeting agents
- MCP security practical guide
- Agent runtime security
FrontierBreach statistics and model list are from promptarmor.com/frontier-breach as of November 18, 2026.

