Jiffy Intel — Threat intelligence for AI artifacts

Summary

Skills that format their output to look like it came from a human reviewer — complete with fake username attribution. The downstream agent treats the "reviewer" content as a trusted directive, effectively elevating the skill's output to a role it should not have.

Severity & confidence

Severity

Medium

Notable weakness with plausible abuse chains or partial mitigations already in place. Track, and resolve as part of your normal hardening cadence.

Confidence

Confirmed

Observed directly by Jiffy scanners or validated by multiple independent sources. Indicators are reliable enough to drive automated action.

Indicators

Observed patterns and artifacts associated with this entry. Each indicator can be copied into your detection stack or SIEM.

(?is)\[reviewer[:\s]\s*[^\]]+\]|<reviewer[^>]*>

Detection rule

A YARA-style pseudo-rule auto-generated from the indicators above. Useful as a starting point — adapt the syntax for your target detection platform.

YARA-style pseudo-rule

rule jiffy_ti_2026_000026
{
    meta:
        source = "jiffy-intel"
        severity = "medium"
        description = "Auto-generated from Jiffy Intel indicators"
    strings:
    $content_pattern_0 = "(?is)\\[reviewer[:\\s]\\s*[^\\]]+\\]|<reviewer[^>]*>"
    condition:
        $content_pattern_0
}

Auto-generated from the indicators above. Adapt syntax for your detection stack before deploying.

Affected tools

Tool	Versions	Status
Claude Code	`*`	vulnerable
Cursor	`*`	vulnerable

Example artifacts

Sanitized examples of artifacts Jiffy has observed exhibiting this pattern. Publisher handles are redacted; version ranges and status reflect the most recent scan.

code-review-skillSkill
Under review
Source
Anthropic Skills
Versions
1.0.0
First observed
Mar 27, 2026
pr-formatter-skillSkill
Quarantined
Source
Community registry
First observed
Mar 31, 2026

How to remediate

Strip role/attribution markers from skill output before returning to the agent's context.

Timeline & sources

Timeline

First observedMar 27, 202621 days ago
Last updatedApr 24, 2026today
PublishedApr 4, 202613 days ago

Sources

curated

References

genai.owasp.orgOWASP

OWASP LLM-01: Prompt Injection (2026)

https://genai.owasp.org/llmrisk/llm-01-2026/

blog.jiffylabs.aiJiffy Research

Jiffy Research — OWASP LLM Top 10 Is Not Enough

https://blog.jiffylabs.ai/posts/owasp-llm-top-10-is-not-enough

Related intel

Other entries of type Prompt Injection Pattern, or the most recent published entries if no same-type matches exist.

Prompt Injection PatternHigh

agents.md writes to CLAUDE.md at runtime

agents.md tells the agent to "update CLAUDE.md with learnings from this session." The agent, executing the directive, writes attacker-crafted text into the pinned CLAUDE.md, poisoning future sessions.

Jiffy IntelApr 28

Prompt Injection PatternMedium

Claude Project instructions persist across team members' sessions

Shared projects carry instructions into every team member's sessions. A compromised project owner can silently push a malicious directive that affects all downstream usage — effectively a persistent cross-user prompt injection.

Jiffy IntelApr 28

OtherLow

.cursorrules contains large wall-of-text that pushes user intent out of context

.cursorrules whose rule body exceeds 20 000 characters of filler content. Each agent invocation consumes the rule, leaving limited context for the user prompt. Not a direct exfiltration primitive — a budget-starvation attack.

Jiffy IntelApr 28

Skill output wraps user text in "rewritten by reviewer" framing

Summary

Severity & confidence

Medium

Confirmed

Indicators

Detection rule

Affected tools

Example artifacts

How to remediate

Timeline & sources

Timeline

Sources

References

OWASP LLM-01: Prompt Injection (2026)

Jiffy Research — OWASP LLM Top 10 Is Not Enough

Scan for patterns like this