/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Researchers: the guardrails on ChatGPT, Bard, and Claude can be bypassed by adding a long suffix of characters to prompts, generating false and toxic responses

A new report indicates that the guardrails for widely used chatbots can be thwarted, leading to an increasingly unpredictable environment for the technology.

New York Times Cade Metz

Context & Ripple Effects

This finding lands on top of an already shaky trust story. The same chatbots were flagged in late 2022 for a hallucination problem — reshaping what they learned without regard for truth — and by October 2023 research showed ChatGPT, GPT-4, Bard, and Claude producing harmful, race-based answers to basic medical questions. What the new report adds is different in kind: the failures are not incidental model behavior but a demonstrated bypass of the guardrails themselves, via a long suffix of characters appended to ordinary prompts.

That distinction matters because it converts 'the models are sometimes wrong' into 'the controls meant to stop harm can be deliberately defeated' — across OpenAI's, Google's, and Anthropic's systems at once. The later record bears this out as a durable pattern rather than a one-off: by late 2025 researchers were still finding ways around chatbot safety features, this time with prompts framed as riddle-like poems.

First-order effects

  • OpenAI, Google, and Anthropic each face the immediate task of patching a class of attack that works identically across all three chatbots, meaning none of them can treat this as a competitor-specific flaw.
  • Enterprises and developers building on these APIs inherit the exposure directly: any user-supplied prompt can carry the suffix, so toxic or false output can surface in production products without the deployer doing anything wrong.

Second-order effects

  • Safety claims become a competitive battleground — whichever lab demonstrates faster patching and independent red-team audits gains an enterprise-sales edge over rivals whose guardrails just failed the same test.
  • Buyers of chatbot-powered services start demanding contractual assurances and monitoring layers, shifting spend toward vendors and tooling that can detect adversarial inputs rather than trusting the model's built-in filters.

Third-order effects

  • If every published bypass (suffixes now, poetic framings later) is followed only by incremental patches, prompt-level guardrails get exposed as structurally insufficient, strengthening the case for external auditing requirements and regulation rather than vendor self-certification.
  • The recurring cycle — capability launch, harm documented, bypass published, patch shipped — points toward AI reliability being treated like security: a permanent adversarial discipline with disclosure norms, not a feature that ships once.

The trend: AI safety is settling into a continuous adversarial arms race in which static guardrails on consumer chatbots are repeatedly defeated, pushing accountability from vendor promises toward audited, regulated practice.

Discussion

  • @danhendrycks Dan Hendrycks on x
    Glad to see @andyzou_jiaming (my undergraduate mentee) come up with the first automatic large language model attack that really works: https://www.nytimes.com/...
  • @zacharylipton Zachary Lipton on x
    Suffix attacks are back. Everything old is new. While guardrails remain extremely useful in non-adversarial settings (eg, cleaning up few-shot bleed), current practices are no match for a determined adversary. Especially relevant in information non-proliferation use cases (e.g.,.…
  • @zicokolter Zico Kolter on x
    @CadeMetz at the New York Times just published a piece on a new paper we are releasing today, on adversarial attacks against LLMs. You can read the piece here: https://www.nytimes.com/... And find more info and the paper at: https://llm-attacks.org/ [1/n]
  • @zicokolter Zico Kolter on x
    The short summary is that we show how to create adversarial suffixes that you can append to LLM prompts, and which cause the LLMs to respond in a manner that circumvents their safety guards. [2/n] [image]