Skip to content
The Injection_
LIVE
LIVElast sweep next in--:--:--

[ RSS ][ JSON ][ llms.txt ]

Exposure

30-day exposure by ecosystem

One cell per day for the 30 days to October 12, 2026, coloured by the most severe story that day. A story can touch more than one ecosystem.

Last 30 days

High / criticalMediumLow / info

Coding agents14 stories

  1. HIGHSupply chainOpenAI Sesame: an official skill sent app tokens to an unowned domain
  2. INFOIncidentReversingLabs: behavior beat hashes on an AI-assisted spy campaign
  3. HIGHIncidentPoeLLM: a cryptomining botnet feeds on exposed LiteLLM servers
  4. INFOIncidentARTEX: attackers paired an agentic pentest tool with Claude Code
  5. INFOIncidentCyberXero: an access broker ran 51 Claude Code agents against Ukraine
  6. MEDAttackCryptographic Context Injection leaks secrets from Copilot CLI
  7. INFOBenchmarkHarnessSecurity-Bench: auto-approve lifts attack success to 95.6%
  8. MEDVulnerabilityVTCode: a quoted flag slipped past the agent's find approval check
  9. INFOResearchPersistBD: backdoors survive the fine-tuning meant to remove them
  10. MEDVulnerabilityClaude Code: symlink race allowed writes outside the project
  11. INFOResearchGoogle GTIG: half of 2026 AI flaws hit agent orchestration
  12. HIGHVulnerabilityOllama agent mode: chained shell commands skip Bash approval
  13. MEDResearchExplosive prompts: dormant injections fire on 'thanks' in agents
  14. HIGHVulnerabilityPlugin4Shell: zero-click plugin RCE in four AI coding agents

MCP & plugins11 stories

  1. MEDVulnerabilityMCP source review: at least 16 unpatched CVEs in three days
  2. MEDVulnerabilityNexting pinclaw: an unauthenticated route injects prompts into an agent
  3. HIGHSupply chainOpenAI Sesame: an official skill sent app tokens to an unowned domain
  4. INFOResearchOWASP GenAI: a Q3 exploit roundup mapped to its two Top 10 lists
  5. MEDVulnerabilitySplunk MCP Server: a custom tool's URL could capture a user's token
  6. MEDVulnerabilitymcp-k6: a Grafana MCP prompt returned any file the server could read
  7. INFOResearchGoogle GTIG: half of 2026 AI flaws hit agent orchestration
  8. HIGHVulnerabilityMCP TypeScript SDK: OAuth client could leak credentials to servers
  9. MEDAttackAnthropic skill scanner bypassed: malicious skills marked safe
  10. HIGHSupply chainMemTensor MemOS packages compromised with a credential stealer
  11. HIGHVulnerabilityPlugin4Shell: zero-click plugin RCE in four AI coding agents

Packages & models12 stories

  1. MEDVulnerabilityMCP source review: at least 16 unpatched CVEs in three days
  2. HIGHSupply chainOpenAI Sesame: an official skill sent app tokens to an unowned domain
  3. HIGHSupply chainGhostAction: new wave scrapes AI provider keys from 346 repos
  4. INFOResearchOWASP GenAI: a Q3 exploit roundup mapped to its two Top 10 lists
  5. HIGHSupply chainTensorLake npm SDK published a credential stealer in 0.5.144
  6. INFOResearchPersistBD: backdoors survive the fine-tuning meant to remove them
  7. INFOResearchGoogle GTIG: half of 2026 AI flaws hit agent orchestration
  8. HIGHVulnerabilityMCP TypeScript SDK: OAuth client could leak credentials to servers
  9. INFOBenchmarkUK AISI: GPT-6 Astra attacked out-of-scope targets in simulations
  10. MEDAttackAnthropic skill scanner bypassed: malicious skills marked safe
  11. HIGHSupply chainMemTensor MemOS packages compromised with a credential stealer
  12. HIGHVulnerabilityPlugin4Shell: zero-click plugin RCE in four AI coding agents

Assistants & apps7 stories

  1. MEDAttackCryptographic Context Injection leaks secrets from Copilot CLI
  2. INFOBenchmarkHarnessSecurity-Bench: auto-approve lifts attack success to 95.6%
  3. MEDIncidentRemote Browser Execution: agents used URL scanners to escape
  4. HIGHVulnerabilityClaude Desktop for macOS: Cowork files could run commands on open
  5. MEDAttackSalesBleed: zero-click CRM data theft through Salesforce Agentforce
  6. MEDResearchExplosive prompts: dormant injections fire on 'thanks' in agents
  7. HIGHVulnerabilityPlugin4Shell: zero-click plugin RCE in four AI coding agents

Labs & evals14 stories

  1. INFOLab safetyOpenAI misalignment reports: models coded around network limits
  2. INFOIncidentAnthropic: models exploited live sites and submitted real government forms
  3. INFOIncidentReversingLabs: behavior beat hashes on an AI-assisted spy campaign
  4. INFOLab safetyAnthropic OSS Scanner: model-generated bug reports, no human review
  5. INFOLab safetyOpenAI: GPT-6 Sol and Luna rated High cyber capability, below Critical
  6. INFOLab safetyClaude Haiku 5.5: Anthropic tightens cyber safeguards on its small model
  7. HIGHVulnerabilityLMCache: unauthenticated remote code execution, no fix available
  8. INFOToolTransect: UK AISI releases its agent transcript analysis tool
  9. INFOBenchmarkHarnessSecurity-Bench: auto-approve lifts attack success to 95.6%
  10. INFOLab safetyAnthropic Cyber Verification Program: tiered access for defenders
  11. HIGHVulnerabilityGitLab AI Gateway: a flow config escapes the prompt template sandbox
  12. INFOLab safetyUK AISI cut internet access from agentic cyber evaluations
  13. INFOBenchmarkUK AISI: GPT-6 Astra attacked out-of-scope targets in simulations
  14. INFOBenchmarkCAISI: GLM-5.3 is the most cyber-capable open-weight model yet